Content uploaded by Sinem Perk
Author content
All content in this area was uploaded by Sinem Perk
Content may be subject to copyright.
Monitoring, Analysis and Diagnosis of
Distributed Processes with Agent-Based
Systems
Ali C¸inar∗Sinem Perk ∗Fouad Teymour ∗Michael North ∗∗
Eric Tatara ∗∗ Mark Altaweel ∗∗
∗Department of Chemical and Biological Engineering, Illinois Institute
of Technology, Chicago, IL 60616 USA (e-mail: perksin@iit.edu)
∗∗ Argonne National Laboratory, Argonne, IL 60439 USA (e-mail:
north@anl.gov)
Abstract: Multiagent systems provide a powerful framework for developing real-time process
supervision and control systems for distributed and networked processes by automating
adaptability and situation-dependent rearrangement of confidence to specific monitoring and
diagnosis techniques. An agent-based framework for monitoring, analysis, diagnosis, and control
with agent-based systems (MADCABS) is developed and tested by using detailed models
of chemical reactor networks. MADCABS is composed of three main hierarchical layers, the
physical communication layer, the supervision layer and the agent management layer. The
supervision layer consists of agents and methods for data preprocessing, process monitoring,
fault diagnosis, and control. The agent management layer conducts the assessment of agent
performances to assign the priorities for selecting the most useful methods of process supervision
for specific types of situations. The paper illustrates the operation of MADCABS for monitoring
and fault detection.
Keywords: Multi-agent systems, process monitoring, fault detection, fault diagnosis, process
supervision, process control, distributed systems, distributed artificial intelligence,
autocatalytic reactions.
1. INTRODUCTION
Multi-layered and adaptive multiagent systems (MAS)
provide a powerful framework for developing a new gen-
eration of real-time supervision and control systems for
distributed and networked processes. The strategy, tech-
niques and tools are being developed at IIT for monitoring,
analysis, diagnosis, and control with agent-based systems
(MADCABS) that automates knowledge extraction from
data, analysis, and decision making. This distributed ar-
tificial intelligence framework is expected to enable the
consideration of novel configurations for manufacturing
such as distributed reactor networks to produce high-
value-added specialty chemicals.
There are strong reasons for distributing the activities
and intelligence in software for supervision of distributed
process operations:
•The complex layout of a manufacturing process yields
a problem that is physically distributed,
•The supervision problem is distributed and heteroge-
neous in functional terms,
•The complexity of the supervision problem dictates
a local point of view that contributes to the de-
velopment of system-wide decisions that may force
reexamination of local decisions,
This work is supported by the National Science Foundation CTS-
0325378 of the ITR program.
•The supervision system must be able to adapt to
changes in the structure or environment of the su-
pervised process or network.
Agents are capable of acting, communicating with other
agents, perceiving their environment, and determining
behavior to satisfy their objectives. They are endowed
with autonomy and they possess resources. However, the
MAS framework offers challenges as well: Agents have only
partial information about their environments, they may
act “selfishly” or initiate actions that may conflict with
actions of other agents. This may lead to undesirable or
harmful behavior in MADCABS or the supervised process,
compromising its profitability and safety.
The nature of the supervision problem dictates the use of
multiple layers of agents where lower-level agents perform
local well-defined tasks such as information validation from
sensors and higher-level agents perform more global tasks
over wider regions of the supervised system. Several agents
can be used to perform a specific task, each using different
methods to enable not only decision by consensus-building
but also to reduce the influence of the weaker methods over
time. Intelligence and adaptation is provided both at agent
and at system level.
MADCABS is composed of three main hierarchical layers,
the physical communication layer, the process supervision
layer and the agent management layer. The physical layer
is where two-way information communication between the
process and MADCABS takes place. Process informa-
tion, such as the flowchart of the process is mapped to
MADCABS through the physical communication layer.
The supervision layer consists of agents and methods for
data preprocessing, process monitoring, fault diagnosis,
and control. Data preprocessing agents filter the process
data, check for outliers and missing data, and provide
estimates for them. Monitoring agents detect deviations
from normal operation and trigger the fault detection and
diagnosis (FDD) agents. When abnormal process opera-
tion is validated, FDD is carried out using contribution
plots, statistical methods, and process knowledge. Control
agents range from simple local PI controllers to decen-
tralized plant-wide grade transition agents. Some agents
collaborate to help each other, while others work in a
competing manner to satisfy a global objective. The per-
formances of different agents and methods are evaluated
in the topmost agent management layer. The agent man-
agement layer is responsible for selecting the best per-
forming agents for process monitoring, fault detection and
diagnosis, and process control for the current operating
conditions. The assessment of agent performances guide
the priorities assigned to select the most useful methods
of process supervision for specific types of situations.
In this paper, MADCABS modules for monitoring and
fault detection, and the information flow among them is
discussed. The paper focuses on the architecture and func-
tionality of MADCABS, the automated tools for assessing
the success of its various functions, the redundancies in
MADCABS, and the adaptation of MADCABS based on
the current state of process operations. The communica-
tion and cooperation between different MADCABS mod-
ules is demonstrated with case studies using autocatalytic
CSTR networks. The capabilities of MADCABS in detect-
ing and diagnosing various types of faults are shown.
2. PROCESS MONITORING, FAULT DETECTION
AND DIAGNOSIS
2.1 Statistical Process Monitoring Techniques
Multivariate statistical process monitoring (SPM) tech-
niques are used in this study. Multivariate techniques that
extract the correlation among variables based on princi-
pal component analysis (PCA) provide the basic tool for
monitoring continuous processes (Kourti and MacGregor
(1996); Cinar et al. (2007); Jackson (1980)). PCA is a
multivariate projection method that extracts strong cor-
relations among the variables in a data set, and based
on that information defines a new orthogonal coordinate
space where the coordinate axes are the highest variance
directions. For chemical processes, where large highly-
correlated process datasets need to be monitored, singu-
larity problems may arise. PCA is well-suited for reducing
the dimensionality of the data and capturing the essential
information in the data.
For large processes that involve many processing units and
many process variables with different correlation struc-
tures, a single PCA model for the whole process may
not give sufficient explanation about the process behavior,
may provide unreliable information based on many false
and missed alarms, and may have difficulty localizing
the source cause among so many variables when a fault
is detected. Multiblock methods have been proposed in
literature for large processes, where the process can be
separated into meaningful process blocks, to increase the
efficiency and interpretability of the statistical monitoring
model. Algorithms to handle multiple data blocks include
hierarchical PCA (HPCA) and consensus PCA(CPCA)
(Qin et al. (2001); Wangen and Kowalski (1988); West-
erhuis et al. (1998); Wold et al. (1996)). Multiblock algo-
rithms enable monitoring of the process both locally and
globally. The CPCA method is designed for comparing
several blocks of descriptor variables measured on the same
objects (Wold et al. (1996); Westerhuis et al. (1998)).
Dynamic PCA (DPCA) is an extension of conventional
PCA to deal with multivariate process data that is corre-
lated in time, using a time-lag shift method (Ku et al.
(1995)). The flow of action, namely, monitoring, fault
detection and diagnosis, is the same for all SPM method-
ologies independent of which monitoring algorithm is em-
ployed.
2.2 Fault Detection, Monitoring and Diagnosis Framework
in MADCABS
MADCABSiswritteninJava,usingRepastSimphonyas
the agent building platform (ROAD (2005)). Among the
important features of Repast that MADCABS uses are its
object oriented structure, scheduling tools and its built-
in automated Monte Carlo simulation framework. Repast
also allows users to change, add, delete agents in run time.
In Repast Simphony, a context is defined as a container
where the agents reside. There are three contexts for moni-
toring, fault detection and diagnosis agents in MADCABS.
The communications between different agents in these con-
texts are shown in Figure 1. Statistical models are built by
the monitoring agents. The fault detection agents, which
are the monitoring statistics, assign themselves to the fault
detection organizer agents responsible for each subsystem
that is monitored. A fault is flagged when a consensus
is formed among different fault detection agents on the
existence of a fault in the system. The fault flag triggers
the diagnosis agent. Diagnosis agent uses information from
the neighboring fault detection organizers, finds the most
contributing process variables to the fault on the faulty
subsystem, and investigates the potential reasons behind
the fault.
Monitoring Agents. The monitoring agents are PCAS-
tarter, DPCAStarter and MultiblockStarter agents and a
monitoring organizer (Figure 2). After the system reaches
steady state and a sufficient amount of normal operation
data is available, monitoring starters are scheduled to form
models. For all process units, a local statistical model
is built using both PCA and DPCA. In the distributed
framework, this is achieved by creating as many PCAS-
tarters and DPCAStarters as the number of operating
units. Each PCAStarter agent builds a separate PCA
model since the data and the PC number that is used
to build the models are different for each unit. DPCAS-
tarters work identical to PCAStarters. There is only one
MultiblockStarter for the whole process. In this case, the
data blocks consist of data from each operating unit. The
MultiblockStarter forms a single multiblock model, which
Diagnosis Agent
Local and Global Monitoring Agents
Fault Detection Organizer
Fig. 1. Monitoring, fault detection and diagnosis agents in
a four reactor network
Multiblock PCA
PCA
DPCA
Fig. 2. Monitoring agents. The upper figure represents a
four reactor network, which is simplified in the lower
figure.
enables the monitoring of the process both locally and
globally through block statistics and super statistics. Since
there is only one model, the size retained in the model is
the same for each block.
The starter agents are subclasses of MonitoringStarterPar-
ent class. Consequently, each of the starters extend some
of the characteristics from the parent class and they also
have class specific properties. The methods in each starter
such as the buildModel and startPro jection are common
methods inherited from the parent, as well as the user-
specified parameters. The number of principal components
to be retained in the model is a superclass variable and it
is overridden in each child class. Each model generates two
monitoring statistics for each subsystem, a T2statistic and
an SPE statistic. Three monitoring methods generate a
total of six fault detection agents for each subsystem. The
fault detection agents are contained in the fault detection
context.
Fault Detection Agents. The fault detection agents,
which are the T2and SPE statistics for each subsystem,
assign themselves to the fault detection organizer of their
subsystem (Figure 3). The fault detection organizer is
responsible for keeping count of its fault detection agents,
declaring consensus fault, keeping history of the perfor-
mances of different fault detection agents under different
fault scenarios, and in case of a consensus fault decision,
triggering the diagnosis agent.
Fault Detection Organizer
Fault Detection Statistics
PCA SPE
PCA T
2
Multiblock SPE
Multiblock T
2
DPCA SPE
DPCA T
2
Fig. 3. Fault detection agents.
Statistics values and confidence limits are among the class
variables for each fault detection agent. If the value of
the statistic goes outside of the limits, the agent flags a
fault. Fault detection organizer keeps track of all the fault
flags given by its statistics. There are several criteria to
form a consensus among different fault detection agents.
The simplest would be to flag a fault, if the majority of
the fault detection agents are flagging fault. This would
require four of the six agents to flag a fault in order to
declare that there is a fault in the unit. In the following,
this strategy will be referred to as the “number weighted ”
consensus criteria.
Another criterion is based on the performances of fault
detection agents over time, and based on their relia-
bility, their decision is given more weight compared to
less reliable fault detection agents. This is referred to
as the “reliability weighted ” consensus criteria. At each
time point, when a new observation is available and new
monitoring statistics are calculated, fault detection agents
either detect a fault or not. Based on their decisions,
they are given an instantaneous performance reward. The
rewarding strategy is designed so that a missed alarm is
penalized the most and the correct detection of fault is
rewarded the most. A set of instantaneous performance
rewards or penalties is given in Table 1, where the rows
show the consensus and columns show the individual agent
decisions. If the fault detection agent flags a fault but the
consensus decision indicates otherwise, the agent is penal-
ized for a false alarm. If the agent does not flag a fault, but
the consensus flags a fault, then the agent is penalized for
a missed alarm. A missed alarm is considered to be worse
than a false alarm, and this is reflected in the instanta-
neous performance calculations. The instantaneous perfor-
mances are summed in time for each detection agent, and
makes up the accumulated performances. The reliability of
an agent is determined by the accumulated performance
values divided by the total accumulated performance value
of all agents in that unit. The reliability weights are then
considered in the consensus decision making.
Table 1. Instantaneous performance rewards
Not faulty Faulty
Not faulty 0.5 -0.5
Fau l ty -1 1
The challenging problem of SPM methods is the missed
and false alarm rates. For some cases, where the fault is
diffusing in the process and affecting the neighboring units
and also with minor faults, the consensus flag may be oscil-
latory. This oscillation affects the performance mechanism
in an undesired way such that an agent that has been
flagging fault in the oscillatory period may not be reliable
enough at that point to affect the consensus decision, and
it will be penalized for flagging fault although the flag
was right. Or an insensitive method could be rewarded
if it did not flag the fault and again this would affect
the consensus in an undesired way. In order to prevent
these, the performances of agents are updated after fault
episodes. A fault episode starts when a consensus fault
is flagged. And the episode continues until no consensus
fault is flagged for eight consecutive time points. At that
point, looking back in history, the performances of agents
are updated.
In order to design an automated fault detection frame-
work, where the decisions of fault detection and diagnosis
highly influence the succeeding tasks, reliability of the de-
cisions is very important. Some of the monitoring methods
may perform better than the others for various states of
the process. Use of agent-based cooperation between differ-
ent methods that are competing for the same task results
in better overall performance than if these methods were
used independently. Several monitoring agents have been
implemented in MADCABS to provide diversity. The aim
is to design an automated fault detection framework that
can detect the faults on time, and that gives fewer false
and missed alarms than if the monitoring methods were
used independently. Comparison of different combinations
of monitoring methods and the false and missed alarm
rates of the corresponding monitoring statistics indicates
that cooperation among agents improved the false and
missed alarm rates.
Table 2. False and missed alarm summary of
different fault detection agent combinations,
using reliability weight condition
Agent Missed Alarm False Alarm
PCA 1.09 0.17
Multiblock PCA 20.98 0.01
DPCA 0.18 0.02
PCA-Multiblock PCA 1.21 0.06
DPCA-Multiblock PCA 0.75 0.07
PCA-DPCA 0 0.03
ALL 0 0.03
Diagnosis Agent. Diagnosis agent works in an event-
driven way. It is activated when a consensus fault is flagged
by a fault detection organizer. The responsibility of the
diagnosis agent is to investigate the type and severity
of the fault under consideration. Contribution plots are
used as a diagnostic tool. The contributions of process
variables to each fault detection statistic are calculated
for different monitoring methods. The contribution plots
do not indicate the source cause of the fault, but identify
the variables that have contributed to the inflation in the
SPM statistic. The diagnosis agent performs contribution
plot analysis and determines the variables that inflated
the SPM statistic that went out-of-control. At this point
a sequence of events is activated. First, the most con-
tributing variable to each statistic is chosen by checking
if the variable contribution value is beyond the 3-sigma
confidence limits and if it makes up a significant amount
of the contributions. Each fault detection agent identifies
their most contributing variables, and the common top
most contributor to all is identified. That variable is then
eliminated from the monitoring model data matrix and
a new statistical model is built using the remaining vari-
ables. The aim is to detect if the fault is a sensor fault or
a process fault. The assumption that is made here is, that
a process fault usually has a fault signature and is realized
in more than one variable. On the other hand, a sensor
fault, especially a single sensor fault does not affect the
other variables if it is not being used for control.
If the new observation, after the variable is eliminated, is
in-control with the new model, it is declared as a potential
sensor fault, however, the projection onto the new model
continues in parallel to check if it will turn out to be a
process fault later since some of the minor process faults
can be misinterpreted as sensor faults in the beginning. If
the new observation is not in-control with the new model,
this means other variables have also been affected from
the fault, and it is declared as a potential process fault. In
addition to discriminating the types of faults, diagnosis
agent estimates the severity of the fault by looking at
how much the variable contributions have gone outside
the confidence limits, how many of the neighboring fault
detection organizer are signaling fault and how many of
the fault detection statistics in the unit have identified the
same fault signature.
3. MONITORING AND FAULT DETECTION OF A
REACTOR NETWORK
3.1 Autocatalytic CSTR Network
Reactor networks hosting multiple species have a very
complex behavior (Figure 2). As the number of steady
states of the network increases, autocatalytic species are
allowed to exist in the network that would otherwise not
exist in a single CSTR. The cubic autocatalytic reaction
for a single autocatalytic species is
R+2P→3Pand P→D(1)
R is the resource concentration, P is the species concen-
tration, D is a dead species. The reaction rate for the
first reaction, species growth rate constant, is kand kd
is the species death rate constant. The feed flow rates
and interconnection flow rates are treated as manipulated
variables. Each reactor has an inlet and outlet flow. The
resource concentration in each reactor along with species
concentrations is also available.
MADCABS is designed to work with process data from
a real process plant or a process simulator. The data are
stored in a database in MADCABS for use by MADCABS
agents. For the case studies, the data are obtained from a
simulator of the CSTR network, where multiple competing
species coexist in the network and consume the same single
resource. The ordinary differential equations modeling the
operation of the reactors are written in C and connected to
Repast Simphony through a Java Native Interface (JNI).
0 50 100 150
0.032
0.034
0.036
0.038
0 50 100 150
0.9
1
1.1 x 10−5
0 50 100 150
0.12
0.13
0.14
0.15
0 50 100 150
8
8.2
8.4
8.6 x 10−8
0 50 100 150
0.015
0.016
0.017
0 50 100 150
1
1.005
1.01 x 10−3
0 50 100 150
1
1.005
1.01 x 10−3
Fig. 4. A 5% process fault in the top right corner reac-
tor 3.(a) Resource concentration in the reactor, (b)
Species 1 concentration in the reactor, (c) Species 2
concentration in the reactor, (d) Species 3 concentra-
tion in the reactor (e) Feed flow rate into the reactor
(Variable 5), (f) Outflowing interconnection to reactor
2 (g) Outflowing interconnection to reactor 7.
3.2 Monitoring and Fault Detection with MADCABS
The case studies use a four-by-five rectangular CSTR
network, where three species coexist feeding from the same
resource. Faults with different magnitudes and types are
simulated to show the effectiveness of the agent-based
monitoring, fault detection and diagnosis framework in
MADCABS.
Fault Detection and Diagnosis. In Figure 4, a step
decrease in the feed flow rate is introduced to reactor 3.
The process fault affected the host species in the reactor
since they start to die. The resource concentration in the
reactor has increased after a delay after the dominant
species start dying. From the figure, the variables that have
been contributing to the fault are seen in the first three
rows of the first column, the resource concentration in the
reactor, dominant species concentration in the reactor and
the feed flow rate to the reactor, which was reset to its
original value after some time.
The contribution plots are given in Figure 5. For the
five fault detection statistics that detected the existence
of a fault in the system, PCA T2, DPCA statistics and
multiblock SPE showed that the main contributor to the
fault is variable 5. PCA-SPE statistic had a smearing
effect, where the signature of the fault could not be seen.
Therefore, having multiple statistics improved diagnosis
results as well. The fault has been detected on time by
five fault detection agents.
Multiblock T2agent is insensitive to faults with magnitude
less than 10%. A contribution chart that shows how many
fault detection agents detected a contributing variable
(Figure 6) indicates that variable 5, the feed flow rate to
Reactor 3, is the common most contributor.
Table 2 provides a summary of 100 runs for each scenario.
Multiblock PCA suffered from its insensitive T2agent,
which had the highest missed alarm rate. The effect of the
1 2 3 4 5 6 7
0
5
10
15
PCA SPE
1 2 3 4 5 6 7
0
200
400
PCA T2
1 2 3 4 5 6 7
0
200
400
Multiblock SPE
1 2 3 4 5 6 7
0
50
100
DPCA SPE
1 2 3 4 5 6 7
−100
0
100
200
DPCA T2
Fig. 5. Contribution plots for reactor 3 at the time of
detection.
1 2 3 4 5 6 7 8 9 10
0
1
2
3
4
5
Fig. 6. Number of fault detection agents and common
contributors for reactor 3.
insensitive statistic to the performance is realized in the
performance of every combination with multiblock PCA.
Combinations with DPCA improved the false and missed
alarm rate. Especially PCA-DPCA performance is supe-
rior to any of the other less diverse combinations and seems
to have a large impact on the combined performance when
all three are used together. In summary, the results show
that having multiple methods working together improves
the effectiveness of the combined monitoring and fault
detection.
Another type of fault in processes is sensor faults, where
sensors might be defective and may provide false readings.
A sensor fault should be identified in a timely manner
since the measured variable can be used in computing the
control actions and an erroneous reading may move the
process to an undesired state, and may even destabilize
the system. In general, correct and timely diagnosis and
communication between control and diagnosis is required.
A sensor fault in the form of a ramp decrease is given to
the resource concentration sensor, variable 1 in reactor 6
(the figure is not provided because of space limitations).
This sensor fault does not affect the other variables since
it is not used for control.
The contribution plots show that the most contributing
variable to the fault is variable 1. When this variable
is taken out of the statistical model data, and a new
model is built with one less variable, the new model
reveals that the process is in-control. This indicates a
potential sensor fault. The diagnosis results for reactor
6 are shared with control agents and also preprocessing
agents in MADCABS so that preprocessing agents can
provide reliable estimates instead of the faulty sensor,
and control agents will continue to provide the necessary
control actions to continue the desired operation level.
As another fault scenario, four consecutive faults are in-
troduced to reactor 3. The fault is again introduced to the
feed flow, and is of magnitude 1%. The fault introduction
times and the detection times of the best performing
combinations are given in Table 3. The reliability weight
based consensus formation is shown to provide much ear-
lier detection times than the majority based consensus
criteria. Since the adapting reliability weight of the fault
detection agents are taken into consideration, the first
criteria provided earlier detection times, for consecutive
faults. The results showed some kind of a learning pattern.
However, this is going to be tested with different validation
cases, where the training is followed by validation with dif-
ferent fault magnitudes. In Table 2, performance of DPCA-
PCA combination was the same as the ALL combination,
however, in Table 3, the detection times showed that when
ALL of the monitoring agents are used in fault detection
the fault is detected earlier than DPCA-PCA combination.
Table 3. Fault detection times (four consecu-
tive process faults)
Agents Fault#1 Fault#2 Fault#3 Fault#4
Actual Fault Times 220 250 290 330
PCA-DPCA 221.9 250.9 290.4 330.8
(reliability)
ALL(reliability) 221.8 250.6 290.1 330.1
ALL(number) 223.1 254.4 294.0 333.6
The average number of agents that are flagging a fault
when the consensus gives a fault flag is listed in Table 4
where two different consensus forming criteria are com-
pared. When the agents’ reliabilities increase with fault
detection, less agents are required to declare a fault. When
the presence of the majority of the fault detection agents is
required to give a fault flag, the missed alarm rates increase
and detection is delayed.
Table 4. Number of agents that flag fault at
the time of detection (four consecutive process
faults)
Agents Fault#1 Fault#2 Fault#3 Fault#4
ALL(reliability) 4.06 3.46 3.48 3.25
ALL(number) 4.26 4.19 4.13 4.20
When the performances of different monitoring methods
are compared the worst performing method is the multi-
block PCA method because of the insensitive T2statistic.
The overall performance of the PCA method is close to
DPCA, but inferior because of the sensitive SPE statis-
tic that gives many false alarms. The methods and the
statistics are ranked and the results are provided in Table
5.
Table 5. Performances of the monitoring agents
Rank Agent
1 DPCA SPE
2 Multiblock PCA SPE
3PCAT2
4PCASPE
5DPCAT2
6 Multiblock PCA T2
4. SUMMARY AND CONCLUSIONS
PCA, DPCA and multiblock PCA methods are widely
used multivariate SPM methods in process industries.
However, all these methods are prone to false and missed
alarms. The common practice in SPM is to test different
monitoring tools form the literature, improve and tune the
algorithms and find the best method that provides reliable
monitoring. Considering the shortcomings of relying on
a single SPM tool, consensus from several SPM tools
is desirable. This is especially important for distributed
and networked processes. Since there are multiple units, a
monitoring system that is giving frequent false alarms on
different operating units will be misleading and will not
be relied on.
In order to improve the effectiveness of monitoring, sev-
eral monitoring methods have been used together in the
proposed framework. Some of the methods performed well
on minor faults and disturbances, but had problems in
contribution charts. Others gave good diagnostic results.
Combining all these methods improved the effectiveness
of the proposed overall monitoring, fault detection and
diagnosis framework.
MADCABS provides an excellent environment to assess
the performance of various SPM and fault detection meth-
ods for specific regions of process operation and adapt the
reliance to different techniques based on prior experience
and recursive assessment of performances. The agent man-
agement layer offers the tools and metrics to assess the
performance of the monitoring, detection and diagnosis
tools and dynamically update the confidence to specific
techniques in a context-dependent way.
REFERENCES
Cinar, A., Palazoglu,A., and Kayihan, F. (2007). Chemical
Process Performance Evaluation. CRC Press, Boca
Raton, FL.
Jackson, J.E. (1980). Principal components and factor
analysis: Part I-principal components. J Qual Technol,
12, 201–213.
Kourti, T. and MacGregor, J. (1996). Multivariate SPC
methods for process and product monitoring. JQual
Technol, 28, 409–428.
Ku, W., Storer, R.H., and Georgakis, C. (1995). Dis-
turbance detection and isolation by dynamic principal
component analysis. Chemometr Intell Lab, 30, 179–196.
Qin, S., Valle, S., and Piovoso, M. (2001). On unifying
mutliblock analysis with application to decentralized
process monitoring. J Chemometr, 15, 715–742.
ROAD (2005). Repast organization for architec-
ture and design. Repast Simphony. Available at
http://repast.sourceforge.net.
Wangen, L. and Kowalski, B. (1988). A multiblock
partial least squares algorithm for investigating complex
chemical systems. J Chemometr, 3, 3–20.
Westerhuis, J., Kourti, T., and MacGregor, J. (1998).
Analysis of multiblock and hierarchical PCA and PLS
models. J Chemometr, 12, 301–321.
Wold, S., Kettaneh, N., and Tjessem, K. (1996). Hierar-
chical multiblock PLS and PC models for easier model
interpretation and as an alternative to variable selection.
J Chemometr, 10, 463–482.
