The document describes a robust approach for analyzing data from fuel cell stack testing. It addresses challenges in handling large amounts of data from multiple devices. The approach includes: 1) Developing an interface in Excel to automate data handling and analysis across Excel and MATLAB for improved efficiency. 2) Using dynamic time warping to synchronize data sequences and align them based on time for better comparison. 3) Applying data reconciliation to optimally adjust measurements so they obey physical constraints like conservation of voltage sums, improving accuracy of analysis.

1. ISA Transactions 51 (2012) 841–847
Contents lists available at SciVerse ScienceDirect
ISA Transactions
journal homepage: www.elsevier.com/locate/isatrans
A robust data treatment approach for fuel cells system analysis
D. Wang n, Y. Zhen
Institute of Chemical & Engineering Sciences, Agency for Science, Technology and Research, 1 Pesek Road, Jurong Island, Singapore 627833, Singapore
a r t i c l e i n f o
abstract
Article history:
Received 15 June 2011
Received in revised form
14 March 2012
Accepted 23 May 2012
Available online 20 June 2012
This paper describes the implementation of a practical approach for fuel cells system data analysis. A
number of data treatment techniques such as data management and treatment, data synchronization,
and data reconciliation are introduced and discussed in order to solve the issues raised in the practical
case. These techniques are integrated in a software environment which provides user a fast, efficient,
and rational electrochemical investigation. The performance of the approach is illustrated using an
industrial fuel cell stack test system.
& 2012 ISA. Published by Elsevier Ltd. All rights reserved.
Keywords:
Fuel cells testing
Electrochemical analysis
Data handling
Data synchronization
Data reconciliation
1. Introduction
A fuel cells stack is a collection of individual cells electrically
connected in series. In an ideal fuel cells stack, every cell would be
subjected to identical operation conditions and the overall stack
performance would be obtained as the sum of the identical output
of individual cells. However, due to manufacturing variability of
components, stack configuration, and degradation with use, the
individual cells in a real stack will typically show some variation
in performance and so will the stack. Hence, in addition to overall
system performance, individual cell performance needs to be
detected in a fuel cells system. Several cell voltage monitoring
systems have been developed for the fuel cells systems and
battery systems [1,8], where the focus is given either on the
design of cell voltage monitoring system, or the investigation of
cell performance by means of the voltage monitoring.
A cell voltage monitoring system has been designed in our
research group for a fuel cells system measurement. A tapping
method is adopted to establish electrical connection to selected
individual cells in a fuel cell stack to explore the coupling
between adjacent cells. Knowledge of individual cells and power
loss related to the interconnect between adjacent cells (i.e.,
bipolar plate) is thus obtained in combination with electrochemical performance of the overall fuel cell stack.
The analysis of data that are measured by the voltage monitoring system takes an important role in obtaining the useful
information of a fuel cells system. Polarization curve is generally
n
Corresponding author. Tel.: þ65 67963959; fax: þ 65 63166185.
E-mail address: david_wang@ices.a-star.edu.sg (D. Wang).
characterized for a fuel cell stack as it reflects the various sources
of polarization (reduction of voltage from the thermodynamically
reversible level) in a fuel cell. The polarization behavior for each
individual cell provides important clues about the system’s
performance and efficiency.
Even though the voltage monitoring system has been investigated by fuel cell designers, the data analysis in the fuel cells
stack testing still remains a challenge. There are some obstacles
that hinder one to obtain the required information easily and
appropriately. One of the problems is with the ease of data
handling. There is a large amount of data collected from several
measuring devices in the voltage monitoring system, the data are
usually stored in several different files (which can be accessed
with spreadsheet software, for e.g. MS Excel). The data have to be
integrated into a statistical software environment (e.g. Origin) in
order to undertake electrochemical calculation, plotting and
analysis. Before a further electrochemical analysis, the raw data
have to be manually examined, synchronised and then transferred from Excel to Origin. This will be a tedious work and timeconsuming, especially when all the procedures have to be
repeated for the new data in the cycle of experiment.
Another challenge is that electrochemical analysis of the data
that are available after the above processing may not produce
reasonable results when comparing the performance between
overall fuel cells stack and individual cells. There are two critical
phenomena that support this argument. It is found that an event
(e.g. voltage change) happened in the experiment is usually
recorded at different time instant and with different logging
points by different testing devices (names of which are mentioned in the following section). It is difficult to synchronize the
data so that the discrepancy in time presents a random fashion
0019-0578/$ - see front matter & 2012 ISA. Published by Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.isatra.2012.05.005

2. 842
D. Wang, Y. Zhen / ISA Transactions 51 (2012) 841–847
Nomenclature
Q,C
time series
m, n
number of points in time series
qi,ci
ith elements/points of time series
d(qi,cj) distance d(qi,cj) between the two points qi and cj
P
a warping path
pk ¼ (i,j)k kth element of warping path
DTW(Q,C) warping cost for two time series Q,C
g(i,j)
the cumulative distance in dynamic programming
x
estimation
(not a regular fashion—for that case a constant time shift of data
could be enough) as shown in Fig. 1, where two data profiles are
horizontally compared. Such a time discrepancy complicates the
analysis and (if not resolved) makes the reasonable analysis
difficult. Another critical phenomenon that has been found in
the investigation of raw data is that the data recorded do not
appear sound in term of the first principles. The relationship
between the voltage of the whole stack and the voltages of
individual cell should satisfy some kind of conservative law, e.g.,
overall voltage equals to sum of individuals as one would expect
in an ideal fuel cell stack. However, this constraint is usually not
satisfied as shown in Fig. 1, where two data profiles are compared
vertically. Also, there is no consistency of the discrepancy but it
appears random. Apparently, directly applying electrochemical
investigation methods to the raw data, even in a user-friendly
handling environment, will hardly give reasonable results.
In this paper, a robust and fast data treatment approach is
developed in order to practically process experimental data
obtained from the voltage monitoring system and to analyze the
complicated fuel cell performance. The approach includes data
automation, data synchronization and data reconciliation.
Firstly, the focus is given on the alleviation of tedium of data
handling cross two software environments. To do this, an
32
Measurement of whole stack
Sum of individual cells
A shift of Sum of individual cells
Design point Start/End by data logger 1
30
Voltage (V)
28
26
24
22
20
18
16
0
200
400
600
800
1000
1200
1400
^
x
estimation
y
measurement
W ¼ diagða2 , a2 ,. . .a2 ,Þ weight matrix with diagonal elements
n
1 2
a2 ,. . .a2
n
1
ai
standard deviation of ith measurement noise
DAE
differential and algebraic equation
A
coefficient matrix of liner equation
Voutput
voltage measurement of whole stack
V cell
voltage measurement of ith individual cell
i
V int
voltage measurement of jth individual interconnect
j
V sic
voltage measurement of kth secondary interconnect
k
interface for data analysis is developed in an Excel file. This
interface includes the functionality with which one cannot only
configure and process the data, but can also take electrochemical
investigation. A computing engine (e.g. MATLAB) is connected
with Excel spreadsheet at background where various algorithms
of data processing and analysis run. It will not affect usage even
though one has little knowledge of MATLAB. This data automation
scheme saves much effort and improves efficiency for data
handling in fuel cell testing, as it is much convenient for one to
undertake analysis in a single software environment where all the
data from different testing devices can be directly accessed and
analyzed.
In order to get the best estimate of data for electrochemical
investigation, the raw data need to be further processed. It is
desirable to identify the similarity among all the data sequences
recorded in the individual testing equipments, and then align the
data in term of a same time instant. (Fig. 1 shows a comparison of
voltage changes measured by different testing devices. It can be
found that the two data sequences have the approximately same
overall component shapes, but these shapes do not line-up in
X-axis (time). The steps (ramping) do not happen at same time,
nor do they follow a consistent fashion.) Dynamic time warping
(DTW) can be used to achieve a better alignment, as it is a method
that allows a computer to find an optimal match between two
given sequences (e.g. time series) with certain restrictions. It has
been successfully used in speech recognition, gesture recognition,
alignment of batch profiles and medicine [5,2,3,6,10]. Its application to fuel cell system data analysis will be a novel approach.
Another concern is to estimate the ‘‘real’’ values of the
measurements from all the devices subject to the first principles.
The vertical discrepancy between the two sets of data as shown in
Fig. 1 is hardly acceptable and less meaningful in terms of the first
principle; at least, the law of conservation should not be breached. Data reconciliation is a procedure of optimally adjusting
measured data so that the adjusted values obey the conservation
law and other constraints. Although data reconciliation has been
applied to chemical processes for process optimization, monitoring and control [4,7], its application to fuel cells system testing
has not been reported.
2. Experimental platform
Sample time
Fig. 1. The fuel cells stack voltages obtained by combining data from several
different measuring devices during polarization curve scanning. The dash line
represents the whole stack voltages measured by one device; the solid line
represents the sum of individual cell voltages that is obtained by combining data
from the other two measuring devices; the solid line with cross marker represents
the shift of the solid line. The circle sign represents the specified point recorded by
data logger file. This figure gives an example of the problems that are to be focused
in the work. (For interpretation of the references to color in this figure, the reader
is referred to the web version of this article.)
The fuel cell stack tested in this study contained 30 cells
connected in series. The test platform for electrochemical performance measurement of the fuel cell stack consists of a customized 500 W fuel cell test station and a TDI MCL488 electronic
load. For individual cell voltage monitoring, conductive wires are
connected to the anode or cathode of selected adjacent cells of
the stack, which enables the voltage measurement of individual
cell and power loss related to the interconnect. The cells’ and

3. D. Wang, Y. Zhen / ISA Transactions 51 (2012) 841–847
843
Fig. 2. Layout of multi-tapped fuel cell and the sections where voltages are measured. Around 30 individual fuel cells are connected in series. Voltages generated across
each cell or some adjacent cells are measured. Voltages over the interconnects are measured. The overall output voltage is measured. These data are measured by three
devices and stored in three data files respectively.
interconnects’ voltages are recorded by two multi-channel data
loggers (Giant, UK), which are controlled by two separate computers, respectively. The schematic diagram of the fuel cell stack
testing system in this study is shown in Fig. 2. The overall fuel cell
stack performance is generally conducted in galvanostatic mode
and the testing current is reached with the use of the electronic
load system. The polarization curve is performed from open
circuit to 2.0 A with increment of 0.1 A/step and holding time of
40 s/step. The inlet hydrogen gas with 3% relative humidity flow
rate is maintained at 4 l/min and air flow is maintained at 100 l/
min regardless of current density. The typical polarization curve
of the overall fuel cell stack is presented as red dotted line in
Fig. 1. For easiness of data handling, the voltage data at current of
0.9 A will be focused on in this study, where the corresponding
cell voltage is among 0.7–0.8 V, which is the typical working
potential of fuel cells.
Fig. 3. Integration between MS Excel and MATLAB. User can make options in Excel
interface to let MATLAB to undertake various data treatment, calculation and
plotting.
3. Interface for data automation and manipulation
In fuel cell stack testing, it is desirable that the measurements
could be easily accessed, pre-processed and manipulated, and the
important parameters could be obtained and displayed promptly
whenever they are needed for investigation. Given that there are
three data files generated by three different devices within each
run in the experiment, manually handling these data across MS
Excel and Origin will exert much burden. In addition, much effort
needs to be put on the treatment of these raw data. Most often,
the data have to be pre-processed, aligned and adjusted in order
to make them right for use. These procedures are either tedious or
hardly completed using manual manipulation, needless to say
that they have to be repeated for the new data set in analysis. This
definitely needs a user friendly environment for an easy and fast
data handling.
Since the data files generated by the devices can be accessed
using MS Excel spreadsheet, it would be more convenient to
manage, to analysis and to display the data in this environment.
Considering the various approaches for data treatment as well as
electrochemical computation that are desired, an interface is
developed in Excel for various data handling, treatment, and
electrochemical computation. This interface connects with
MATLAB at background as computational and graphics engine
and it brings the results to display. Spreadsheet Link EX connects
Excel spreadsheet software with the MATLAB workspace,
enabling one to access the MATLAB environment from an Excel
spreadsheet [9]. With Spreadsheet Link EX software, one can
exchange data between MATLAB and Excel, taking advantage of
the familiar Excel interface without leaving the Excel environment while accessing the computational functionality, graphic
interface and visualization capabilities of MATLAB (Fig. 3).
Fig. 4. Main tasks listed in Excel interface for data treatment and analysis. It
displays on spreadsheet with user information for configuration, which allows
user to make options and display the results.
Given the environment, a standard data processing procedure
can be listed in the spreadsheet with brief information on
operation for a user. Correspondingly, at each step of procedure
the programs are edited that aims to undertake a specific task.
User can make options to configure a need in the procedure. Fig. 4
displays an example of spreadsheet interface that includes several

4. 844
D. Wang, Y. Zhen / ISA Transactions 51 (2012) 841–847
main tasks, where user can work on each cell (denoted by 0) for a
specific operation.
32
Measurement of whole stack
Consolidation of indiv. cells, and interconnects
Design point from whole stack measurement
Design point from consolidation
31
30
4. Data pre-processing and synchronization
29
28
27
26
25
24
23
0
50
100
150
200 250 300
Sample time
350
450
500
26
25.8
25.6
25.4
25.2
25
Measurement of whole stack
Condolidation of indiv. cells, and interconnects
Design point from whole stack measurement
Design point from consolidation
24.8
340
350
360
370
380
390
400
410
420
Sample time
Fig. 6. A zoom-in of two data sequences in a sample range.
To align two sequences using DTW one constructs a m  n
matrix where the (ith, jth) element of the matrix contains the
distance d(qi,cj) between the two points qi and cj (typically the
Euclidean distance is used, i.e., d(qi,cj)¼(qi Àcj)2), each matrix
element (i,j) corresponds to the alignment between the points qi
and cj. A warping path P is a contiguous set of matrix elements
that defines a mapping between Q and C. The kth element of P is
defined as pk ¼(i,j)k so one has
P ¼ p1 ,p2 ,. . .pk ,. . .,pK maxðm,nÞ r K om þ nÀ1
4.2. Data synchronization
400
Fig. 5. Two data profiles after the pre-treatment.
Voltage (V)
The quality of data is very important in fuel cell system electrochemical investigation. Suitable pre-treatment of raw data is sometimes crucial for the analysis result. In this study the testing data
offer a unique challenge not only in terms of data handling, but also
in terms of data quality. Missing data points are not uncommon in
measuring device, which make a data file incomplete and its samples
wrongly associated with the ones recorded in other data files. There
are outliers in the data due to equipment malfunction, and there are
no numerical data in the records due to equipment calibration. In
addition, it is necessary that the samples from different files need to
be aligned to describe the phenomena (e.g. voltage change) that
happen at same time instant.
These data quality issues need to be carefully addressed prior
to further electrochemical analysis. Generally, data are checked
visually at first. Anything that appears suspicious from electrochemical point of view is double-checked and carefully considered to determine what treatment is needed. For example, there is
a missing record every 10 s in one data file. These missing data
need to be reconstructed using interpolation approach, which can
be realized by making a choice on the interface and running
functions in the background. The interpolation can also be used to
reconstruct the missing data due to device calibration. Outliers,
which can be simply regarded as the data points that are not
consistent with the bulk of data, are common in experiment data
set. In the univariate approach, outliers are detected based on
visualization together with numerical comparison with adjacent
data. The suspicious outliers can be replaced with median values
of the adjacent data in the approach.
Another important pre-treatment is to align the data from
three files to describe the phenomena that happen at same time.
Without such treatment, the data could not be used for electrochemical investigation with confidence. It is observed in Fig. 1
that the plot of raw data clearly misrepresent the real phenomena, because the step changes do not happen at same time
instant. In order to alleviate this, the variable measurement that
is most sensitive to such step change and trustworthy is selected
as the benchmark for such alignment based on a prior knowledge.
The samples of other variable measurements are shifted accordingly to align with this data set.
The plot of data after above pre-treatment is shown in Fig. 5as
an example. The two profiles are approximately coinciding in
general, but still a discrepancy exists. A closer look at current of
0.9 A as shown in Fig. 6 demonstrates the discrepancy in both
X-axis (sample) and Y-axis (value) although the above treatment
is implemented.
Voltage (V)
4.1. Data pre-processing
ð2Þ
The warping path is typically subject to several constraints:
A better alignment in sample time can be achieved by using
dynamic time warping (DTW) technique, which is used to find the
similarity between time series that have approximately the same
overall component shapes but these shapes do not line up in Xaxis. The following is a brief description on DTW algorithm.
For two time series Q and C, of length m and n respectively,
where:
Q ¼ q1 ,q2 ,. . .qi ,. . .,qm
C ¼ c1 ,c2 ,. . .ci ,. . .,cn
ð1Þ
1. Boundary conditions: w1 ¼(1,1) and pK ¼(m,n). This required the
warping path to start and finish in diagonally opposite corner
cells of the matrix
2. Continuity: given pk ¼(i,j) then pk À 1 ¼(i’,j’) where iÀ i’ r1 and
j Àj’r1. This restricts the allowable steps in the warping path
to adjacent cells.
3. Monotonicity: given the pk ¼(i,j) then pk À 1 ¼ (i0 ,j0 ) where
i À i’Z0 and j À j0 Z0. This forces the points in P to be monotonically spaced in time.

5. D. Wang, Y. Zhen / ISA Transactions 51 (2012) 841–847
There are many warping paths that satisfy the above conditions; however, one is interested in the path which minimizes the
warping cost:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u K
uX
ð3Þ
pk =K
DTWðQ ,CÞ ¼ mint
and the warped data are aligned in X-axis. This indicates that the
data from different measuring devices are being synchronized to
describe the phenomena that happen at the same sample time.
The warping path and matrix of distance are depicted in Fig. 8,
where the distances between the data, the warping path, and two
data sequences are presented. It is worthwhile to note that DTW
may increase the number of the warped data comparing with the
original ones, as DTW tries to map the similarity between the two
sequences so that one point in a sequence may map with several
points in the other. Fig. 9 overlays the warped data on the original
ones. It can be seen that there are four more data point in both
warped data sets. These added data do not affect the investigation
and can be ignored if the data section at specified point is focused.
It is also noticed that the warping usually happen where there is a
step change. This advises that in the electrochemical calculation
that follows, it would be better to select the data that is located in
the middle section of the steps (e.g. the first and the last three
samples of the data in a step can be ignored).
k¼1
This path can be found very efficiently using dynamic programming to evaluate the following recurrence which defines
the cumulative distance g(i,j) as the distance d(i,j) found in the
current cell and the minimum of the cumulative distance of the
adjacent elements:
gði,jÞ ¼ dðqi ,cj Þ þ mingðiÀ1,jÀ1Þ, gðiÀ1,jÞ, gði,jÀ1Þ
845
ð4Þ
Given the data section in Fig. 6, applying DTW to the data gives
the warped data as shown in Fig. 7, where the original is shown in
the left part of the figure, the warped data is shown in the right
part. It can be seen that the pessimistic dissimilarity is eliminated,
Original data
Warped data
26
Output V
Calculated V
25.8
25.8
25.6
25.6
Voltage (V)
Voltage (V)
Output V
Claculated V
25.4
25.4
25.2
25.2
25
25
24.8
24.8
20
40
60
80
20
40
60
80
Samples
Samples
Fig. 7. Comparison of original data and the warped data.
0
74
10
63
20
52
42
40
50
31
60
Distance
Samples
30
21
70
10
80
10
20
30
40
50
60
70
0
V
28
26
24
24 26 28
V
80
Samples
Fig. 8. Distance matrix of DTW, where the two data sequences are shown in the left and underneath. The rightmost bar indicates the distance value. The diagonal curve is
the warping path that maps the data in the two sequences.

6. 846
D. Wang, Y. Zhen / ISA Transactions 51 (2012) 841–847
26
26
Cal. V
Out. V
Cal. V
Out
25.6
Out. V
Voltage (V)
25.8
25.6
Voltage (V)
25.8
Cal
Out
Cal
25.4
25.2
Cal. V
25.4
25.2
25
25
24.8
24.8
24.6
24.6
0
10
20
30
40
50
60
70
80
90
10
20
30
40
50
60
70
80
90
Sample
Sample
Fig. 9. The two original data sequences and the two warped data sequences are
superimposed in one figure. The warping happens at the step changes. More data
points are included in warped data.
0
Fig. 10. Data reconciliation of the measurement data. The pentagrams represent
the reconciled data for both data sequences that are adjusted to be equal in
quantity.
1
5. Data reconciliation
x
s:t:
DAE model and inequality constraints
ð5Þ
where e¼yÀ x, f(e) is the probability density function of the error.
If the sensor errors are assumed to follow Normal distributions,
the objective function Àlog f ðeÞ is quadratic and the conventional
weighted least squares formulation is obtained,
min ðyÀxÞT W À1 ðyÀxÞ
x
s:t:
ð6Þ
0.9
0.8
Voltage (V)
DTW tries to explain the variability in Y-axis by warping the
X-axis so that two data sequences can be aligned in terms of
sample time. However, DTW could not make the data aligned in
terms of quantity. As can be seen in Figs. 7 and 9, there is still
discrepancy in value between the two data sequences before and
after DTW applies, which conflicts the energy balance. It is
necessary to eliminate the errors before the data is rationally
used for electrochemical investigation.
Data reconciliation is a procedure of optimally adjusting
measured data so that the adjusted values obey the conservation
laws and other constraints. The essence of data reconciliation is
that: given the measurement vector y, one wants to estimate its
state x, which satisfies the models from first principles. This
problem can be approached with the laws of probability and
maximum likelihood principle by minimizing the negative logarithm of the probability function of the difference between
measurement y and estimation x, subject to the constraints [11],
_ ¼ argmin½Àlogf ðeÞŠ
x
0.7
0.6
0.5
cell voltage using the approach
cell voltage without using the approach
0.4
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Cell position
Fig. 11. Typical graph for cell voltage vs. cell position of fuel cell stack. It shows
that the voltages obtained without using the approach are out of the reasonable
range (0.7–0.8 V), while the voltages obtained using the approach are within the
reasonable range. The values appear the same among the adjacent cells are the
averages since the voltages across the adjacent cells are measured.
where Voutput is the stack voltage measurement, V cell is individual
i
cell voltage measurement, V int is the measurement for individual
j
interconnect, V sic is the measurement for secondary interconnect.
k
Re-arranging the constraint equation and choosing a set of
suitable standard deviations for the measurement, the estimated
data can be obtained as shown in Fig. 10, where the reconciled
data for both profiles are coincided.
DAE model and inequality constraints
where W ¼ diagða2 , a2 ,. . .a2 ,Þ and ai is the measurement noise
n
1 2
standard deviation. Furthermore, for linear equality constraints
Ax ¼0 and in the absence of inequality constraints, this problem
has a closed-form solution via Lagrange-multipliers,
^
x ¼ yÀWAT ðAWAT ÞÀ1 Ay
ð7Þ
In fuel cell stack testing problem, given the measurement data
from three devices, one wants to estimate their ‘‘real’’ values
subject to the following energy balance:
V output ¼
30
X
i¼1
V cell À
i
29
X
j¼1
V int À
j
2
X
k¼1
V sic
k
ð8Þ
6. Fuel cell system performance analysis
Given the data being processed, the electrochemical calculations can be implemented within the interface so that various
plots can be provided for visualization and analysis. Fig. 11 gives
an example of plot of some parameters. From the individual cell
I–V–P curves data, take the average of the voltage values for each
individual cell obtained when the current is 0.9 A. The voltage
values obtained at this current value are the voltages at specified
point current. The voltage values are tabled against cell positions
on tube. For the cells that are connected and measured as a whole,
an average will be taken for each cell. Plot the voltage values

7. D. Wang, Y. Zhen / ISA Transactions 51 (2012) 841–847
versus the cell positions on tube, and the voltage distribution at
the specified point current (0.9 A) against cell position on tube
will be obtained. Fig. 11 shows the voltage distribution generated
by each cell at current of 0.9 A (denoted by the stars). A
comparison is given by superposing the data that were obtained
initially when the proposed treatment was not applied (denoted
by circles). It can be seen that the previous voltages are higher
than that with the treatment, and they are out of the reasonable
range (0.7–0.8 V).
Many other data analysis and plots can be produced with the
developed approach, but they will not be displayed for the
brevity. With this user-friendly interface and proposed data
treatment, one can undertake a fast, efficient, and rational
electrochemical investigation of fuel cell stack testing.
7. Conclusions
Data handling and processing are ubiquitously topics in
virtually every scientific discipline and industrial practice. While
the basic techniques are by now well known, their sensible
selection, integration and deployment in a real problem still
remains a challenge. Before undertaking electrochemical analysis
in industrial setting, there is always a need to process the raw
data in order to make the analysis either efficient or rational,
or both.
The issues to be considered in this practical case also imply
that there is no single technique that can solve the practical data
analysis problem. It needs a number of techniques, well selected
and seamlessly integrated into a software tool. Even though the
techniques being employed in the approach are not originally
created in this work, they are judiciously selected and they are
firstly applied and customized in this practical case to solve the
issues.
Even though this work presents a solution, there still is a room
to improve. The interface needs to be upgraded to facilitate its
847
usage, the precision in synchronization and reconciliation needs
to be improved. For example, the non-measured interconnects are
not considered in data reconciliation, which may have an impact
on the analysis. This could be eliminated either by an estimation
of these interconnects, or a re-design of the measurement layout.
In addition, suggestion can be given on the current system
architecture that a single computer connecting the data loggers
would minimize the synchronization problem. The authors intend
to pursue these concerns in the future.
References
[1] Brunner D, Prasad AK, Advani SG, Peticolas BW. A robust cell voltage
monitoring system for analysis and diagnosis of fuel cell or battery systems.
Journal of Power Sources 2010;195(24):8006–12.
[2] Godin C, Lockwood P. DTW schemes for continuous speech recognition: a
unified view. Computers Speech and Language 1989;3:169–98.
[3] Kassidas A, MacGregor John F, Paul ATaylor. Synchronization of batch
trajectories using dynamic time warping. AIChE Journal 1998;44(4):864–75.
´
[4] Prata Diego Martinez, Schwaaba Marcio, Luis Lima Enrique, Carlos Pinto Jose.
Nonlinear dynamic data reconciliation and parameter estimation through
particle swarm optimization: application for an industrial polypropylene
reactor. Chemical Engineering Science 2009;64(18):3953–67.
[5] Rabiner LR, Juang B. Fundamentals of speech recognition. Prentice-Hall, Inc.;
1993.
[6] Ramaker Henk-Jan, van Sprang Eric NM, Westerhuis Johan A, Smilde Age K.
Dynamic time warping of spectroscopic BATCH data. Analytica Chimica Acta
2003;498:133–53.
[7] Ramamurthi Y, Sistu PB, Bequette BW. Control-relevant dynamic data
reconciliation and parameter estimation. Computers & Chemical Engineering
1993;17(1):41–59.
[8] Santis M, Freunberger SA, Papra M, Wokaun A, Buchi FN. Experimental
investigation of coupling phenomena in polymer electrolyte fuel cell stacks.
Journal of Power Sources 2006;161:1076–83.
[9] Spreadsheet Link EX, /http://www.mathworks.com/products/excellink/S.
2010.
[10] Tormene P, Giorgino T, Quaglini S, Stefanelli M. Matching incomplete time
series with dynamic time warping: an algorithm and an application to poststroke rehabilitation. Artificial Intelligence in Medicine 2009;45(1):11–34.
[11] Wang D, Romagnoli JA. A framework for robust data reconciliation based on a
generalized objective function. Industrial & Engineering Chemistry Research
2003;42:3075–84.