-
Causal analysis for alarm flood reduction
Vicent Rodrigo ∗,∗∗ Moncef Chioua ∗∗ Tore Hagglund ∗
Martin Hollender ∗∗
∗Department of Automatic Control, Lund University, SE-221 00
Lund,Sweden (email: [email protected],
[email protected]).∗∗ABB Corporate Research Germany, Wallstadter
Strasse 59 (email:
[email protected], [email protected])
Abstract: The introduction of distributed control systems and
the high level of interconnectiv-ity of modern process plants has
caused alarm flooding to become one of the main problems inalarm
management of process plants. A reduction of alarm flood periods
contributes to a decreasein plant incidents. In this work, a
combination of alarm log, process data and connectivityanalysis is
used to isolate consequence alarms originating from the same
process abnormalityand to provide a causal alarm suggestion. The
effectiveness of the method is illustrated on anindustrial case
study of an ethylene plant, a typical example of a large-scale
industrial system.
Keywords: alarm systems, fault detection and diagnosis,
root-cause, plant-wide disturbance,plant connectivity, approximate
sequence matching, large-scale systems.
1. INTRODUCTION
Alarm systems are crucial elements in the operation ofprocess
plants as they continuously monitor the processand inform the
operators if an undesired process state thatrequires their
assessment or action is reached (IEC-62682,2014).
The introduction of distributed control systems in theprocess
industry has increased the number of alarms peroperator
exponentially (Hollifield and Habibi, 2011). Ad-ditionally,
nowadays, modern process plants present a highlevel of
interconnectivity due to steam recirculation, heatintegration and
use of complex control systems. A dis-turbance affecting the
process plant can therefore spreadthrough its material, energy and
information connectionsand affect all the process variables on the
path. Alarmsassociated to these process variables are triggered and
mayoverload the control room operators who will not be ableto
properly investigate each of these alarms (Wang et al.,2015). This
undesired situation known as alarm floodhas been recognized as a
major cause of most industrialincidents investigated by the US
Chemical Safety Board(Beebe et al., 2007). The ANSI/ISA-18.2 (2009)
definesan alarm flood as the occurrence of more than 10 alarmsper
10 minutes and per operator. In such situation, theoperators might
not be able to keep the process plantwithin safe operation
endangering human lives and theenvironment (Bransby and Jenkinson,
1998).
Henningsen and Kemmerer (1995) pointed out that in atypical
alarm log, a significant part of nuisance alarmsfall in the
following categories: repetitive alarms, standingalarms and
consequence alarms. For removing repetitiveand standing alarms
methods such as the use of filter-ing, time-delay, dead-band and
difference functions areproposed. These techniques succeed in
lowering the num-ber of alarms when the process is running under
nomi-nal conditions. Nevertheless if an abnormality occurs and
propagates through the plant, these methods are unable
tosuppress consequence alarms and to provide a proper di-agnosis.
This task demands to identify all causal relationsbetween alarms
(Hollender and Beuthel, 2007).
Recently, several contributions from academia and indus-try
proposed systematic methods for the grouping of con-sequence
alarms. Event correlation is applied in Higuchiet al. (2009) and
Yang et al. (2012) for this purpose.On the other hand, noticing
that a large portion of thealarms occurring in an alarm flood are
related to eachother and follow a specific pattern, Kabir et al.
(2013)and Cheng et al. (2013) suggest using sequence
patternmatching for grouping consequence alarms. Kabir et al.(2013)
presented a method in which the similarity betweenalarm flood
sequences is determined under the assump-tion that alarms occur in
a consequential manner. Thisassumption, may however not always be
true as the or-der of occurrence of alarms is highly dependent on
theirlimit setting. Postulating that the exact order of
alarmoccurrences is not as important as their proximity in
time,Cheng et al. (2013) proposed an alternative method forpattern
matching of alarm flood sequences based on theSmith-Waterman
algorithm. Time stamps of the alarmspresent in the alarm flood
sequence are used to blur theorder of the alarm sequence when the
alarm time stampsare considered close to each other. Lai and Chen
(2015)proposed an extension of this algorithm to the case
ofaligning multiple alarm flood sequences.
In Schleburg et al. (2013), the authors suggest a methodto group
consequence alarms based on both a topologymodel describing the
process connectivity and a set of rulesbuilt using process
knowledge. In this method, every pairof alarms is checked against
the set of rules and the plantconnectivity to evaluate if two
alarms are related.
While grouping related alarms has a clear benefit inassisting
the operators reacting to an alarm flood episode,
Preprint, 11th IFAC Symposium on Dynamics and Control of Process
Systems,including BiosystemsJune 6-8, 2016. NTNU, Trondheim,
Norway
Copyright © 2016 IFAC 723
-
pointing them to the causal alarm (i.e. the alarm relatedto the
root-cause of the alarm flood) could significantlyreduce their
reaction time.
Causal root-cause analysis for process troubleshooting isthe
topic of several recent contributions investigating thedesign of
systematic methods for abnormality propagationpaths determination
and plant-wide disturbances causalanalysis. Quantitative
data-driven methods detecting di-rectionality between measurements
have emerged, such astime delay analysis (Fan and Deyun, 2012) or
transferentropy, which is used to identify the causality
betweenmeasurements even in absence of observable time
delaysbetween them (Bauer et al., 2007). A combination of
quan-titative process history methods and qualitative modelslike
Signed direct graphs (SDG) (Fan and Deyun, 2012)are proposed in
Thambirajah et al. (2009).
The present work follows the line of thought in Schleburget al.
(2013) but aims at developing a systematic methodfor the isolation
of the causal alarm in an alarm flood. Theproposed approach relies
on the use of multiple processinformation sources: alarm logs,
historical process dataand a topology model of the process. Even
though thedata originating from different information sources
differin their nature, they are related to each other.
By combining alarm log, process data and conectivityanalysis not
only related alarms of an alarm flood can begrouped, but also their
causal alarm can be identified.
2. CAUSAL ANALYSIS FOR ALARM FLOODREDUCTION
This section describes an approach for alarm flood reduc-tion
and causal alarm determination.
The proposed approach relies on two assumptions:
(1) Alarm floods are the result of an abnormality prop-agating
in the process through material, energy andinformation
connections.
(2) If two alarm flood sequences are similar, then theystem from
the same process abnormality.
Regarding the first hypothesis, the aim of this work isto reduce
alarm floods caused by a disturbance spreadingthrough the plant due
to its high level of interconnectivity.As for the second
hypothesis, if two alarm flood sequencesshare a common pattern and
if a large portion of the alarmoccurrences in the alarm floods are
present in this pattern,it is very likely that they originate from
the same processabnormality.
The method consists of five steps. In the first step,
chatter-ing alarms are removed from the alarm log. In the
secondstep, time intervals containing alam floods are identifiedand
alarm flood sequences are built. Similar alarm floodsequences are
clustered in the third step. In the fourthstep, the signals and
assets associated with the analyzedalarms are identified. In the
last step, process data causalanalysis followed by connectivity
validation are performedin order to suggest a causal alarm. The
method’s generalworkflow is depicted in Fig. 1.
In the present work it is assumed that an alarm sequenceis
defined as follows:
Mapalarmstosignalsandassets
Clusteralarmfloodsequences
Alarmfloodsequences
Alarmlog
Iden7fyalarmfloodperiods
Removecha=eringalarms
Alarmlogwithoutcha=eringalarms
Abnormalitytemplatesandcorrespondingcausal
alarms
Isolatethecausalalarm
Alarmfloodsequencesclusteredandan
abnormalitytemplateforeachcluster
Abnormalitytemplatesandalarm,signalandassetsiden7fiersmapped
Topologydata Processdata
Fig. 1. Method’s workflow
A = [a1, a2, ..., aM ]am = (em, tm),m = 1, 2, ...,M and em ∈
Σwhere em is the alarm type, particularly a tag containedin the
alphabet of alarm types Σ, and tm its time stamp.
2.1 Remove chattering alarms
Chattering alarms are defined by the industrial
standardANSI/ISA-18.2 (2009) as “alarms that repeatedly tran-sition
between the alarm state and the normal state ina short period of
time”. Since the present work focuseson analyzing consequence
alarms, chattering alarms areremoved prior to the causal
analysis.
Following Kabir et al. (2013), in order to remove
chatteringalarms, a minimal admissible time interval length is
chosenand if the time elapsed between the occurrences of twoalarms
of the same type is lower than the chosen interval,the second alarm
is removed. The minimal admissible timeinterval length in the
present work is taken as 10 min,based on the definition of alarm
flood.
2.2 Identify alarm flood periods
This step isolates time periods where alarm floods occurand
builds the corresponding alarm flood sequences.
The alarm log is divided into time intervals of 10 min.Intervals
with more alarm occurrences than a specifiedthreshold are
highlighted. The threshold (τ) is chosenbased on the definition of
an alarm flood, i.e. more than10 alarms per 10 min and per
operator.
Duration of alarm floods in process plants ranges from afew
minutes to several hours. In order to identify an alarmflood
regardless of its duration, consecutive intervals withmore alarm
occurrences than the specified threshold aremerged. Since the alarm
log is discretized, it might happen
IFAC DYCOPS-CAB, 2016June 6-8, 2016. NTNU, Trondheim, Norway
724
-
that the beginning or the end of an alarm flood sequence iscut
out, in order not to lose part of the sequence, the timeintervals
before and after the consecutive time intervalscontaining more than
τ alarms are also included.
2.3 Cluster alarm flood sequences
Once the alarm flood time intervals are isolated andthe alarm
flood sequences are built, similar alarm floodsequences are
clustered using sequence pattern matching.
Alarm time stamps are taken into account as the proximityin time
of the alarms is more important than the exact or-der of
occurrence. We use the method described in Chenget al. (2013) based
on a modified Smith-Waterman (MSW)algorithm to blur the order of
the alarms in the sequencewhen they occur closely in time. A
similarity matrix isbuilt using pairwise similarity indices
obtained from theMSW algorithm. Finally, agglomerative hierarchical
clus-tering (AHC) with the average linkage criterion is appliedto
cluster the alarm flood sequences.
One main disadvantage of the MSW algorithm is itsheavy
computational burden. In order to overcome thisproblem a
prefiltering stage is added. A less precise butless time consuming
method is used (Kabir et al., 2013).Combinations of alarm sequences
with low similarity arefiltered out and the MSW index computation
is applied toa subset of alarm sequence combinations with a
significantsimilarity. A pair of alarm sequences is considered
similarif the two alarm sequences have a high ratio of alarmsin
common. If two alarm sequences have few alarms incommon, the
computation of the MSW similarity index isdiscarded. A zero
similarity index will be assumed.
Two thresholds are used: the prefilter threshold (λpref )and the
clustering threshold (λclus). The first thresholddecides whether a
pair of alarm sequences is filtered out.The second threshold sets
the stopping criteria of theclustering algorithm. The values of
these two thresholdsare interrelated: for a given clustering
threshold, settingthe prefilter threshold too high can lead to
combinations ofalarm sequences with MSW similarity index large
enoughfor being clustered, but with a prefilter similarity indextoo
low to pass the prefilter. A solution to facilitate thethresholds’
choice is to make the prefilter and the MSWsimilarity indices
comparable. If one could insure thatSIprefilter ≥ SIMSW , for a
given clustering threshold ifthe prefilter threshold is set to be
equal or smaller than theclustering threshold, all pairs of alarms
that have a MSWsimilarity index high enough to be clustered
together willpass the prefilter.
Given two alarm sequences A and B, a similarity indexthat meet
this requirement is:
SIprefilter(A,B) =|A ∩B|
max(|A|, |B|)(1)
The numerator represents the number of common alarmoccurrences
in both sequences. And the denominator is thenumber of alarm
occurrences of the longest sequence (thesimilarity index is
normalized between 0 and 1).
According to assumption (2) in Section 2, sequences be-longing
to the same cluster stem from the same process
abnormality. Therefore, an alarm template can be builtfor each
cluster/process abnormality. This alarm templatecontains the
representative alarms that identify the corre-sponding process
abnormality (Bouchair et al., 2013), seeFig. 1. Alarms that are not
related to a process abnormal-ity can trigger during the same time
interval. Obviously,these alarms should not be included in the
abnormalitytemplate. Filtering out these alarms is achieved based
ontheir frequency of occurrence in the training set of alarmfloods.
The threshold on frequency of occurrence is set to50%, i.e. an
alarm is kept only if it occurs in more than50% of the sequences
belonging to a given cluster.
Set of alarm flood sequences
Pre-filter Similarity indices
Clustering and create templates
Reduced set of combinations of alarm flood sequences
Similarity indices Alarm flood sequences clustered and a fault
template for each cluster
Fig. 2. Clustering alarm flood sequences step
The third step of the proposed method where alarm floodsequences
are clustered and a set of templates definingthe process
abnormalities of all clusters is generated issummarized in Fig. 2
and described in Alg. 1.
Alg. 1: Cluster alarm flood sequencesinput : Alarm flood
sequences {Ai} : i = 1, 2, ..., Noutput: Clusters {Ck} : k = 1, 2,
..., K Ck = {Ai : i ∈ 1, 2, ..., N} and
fault templates {Tk} : k = 1, 2, ..., K Tk = {e : e ∈ Σ}
1 For each pair {Ai, Aj} : i 6= j and i, j = 1, 2, ...,
Nif(SIpref(Ai, Aj) ≥ λpref ) add to list LMSW
2 Build an N ×N similarity matrix S
si,j =
{SIMSW (Ai, Aj), if {Ai, Aj} ∈ LMSW0, otherwise
3 {Ck} = AHC(S, λclus)4 For each Ck if (alarm type e ∈ Σ is
present in more than 50% ofAi ∈ Ck) add e to Tk
2.4 Map alarms to signals and assets
The absence of a common naming convention for alarm,signal and
assets tags hinders a straightforward mappingbetween the different
process information sources. Wepropose to map alarms, signals and
assets based on thesimilarity of the symbols contained in their tag
names.For instance, the tag name of a temperature controllerfrom
the case study is TC11 in the P&ID, this controllerhas a signal
associated whose tag is TC11CO.XAY andan alarm type TC11CO is
assigned to this signal.
In this work six hashmaps {Mapx→y : x 6= y, x, y =e, si, as}
mapping alarm types (e), signals (si) and assets(as) are generated
based on the computation of a similarityindex as described in Alg.
2. Particularly, the similarityindex (SISW ) obtained from the
original Smith Watermanalgorithm (Smith and Waterman, 1981).
Alg. 2: Mapping signals and alarmsinput : Sets of all signal
tags Si={sitag} and all alarm types Σoutput: Mapsi→e and
Mape→si
1 For each sitag , find emap : emap ∈ Σ, SISW (sitag, emap)
=max({SI(sitag, e)}) e ∈ Σ, SISW (sitag, emap) ≥ λmap
2 Add pair [sitag, emap] to Mapsi→e and [emap, sitag ] to
Mape→si
IFAC DYCOPS-CAB, 2016June 6-8, 2016. NTNU, Trondheim, Norway
725
-
Fig. 3. P&ID of the cracking and quenching section of the
ethylene plant
2.5 Isolate the causal alarm
The first alarm with respect to the time of occurrencecannot be
considered as the causal alarm since the timeat which an alarm
triggers is highly dependent on the set-tings of the alarm limits.
Moreover, the time between theoccurrence of a process abnormality
and the correspondingalarm triggering is probabilistic. As the
causal relation-ships between signals can be captured, the process
dataassociated to these alarms should be analyzed instead.
The method used for this purpose is transfer entropy(Bauer et
al., 2007). Transfer entropy uses transition prob-abilities to
estimate the amount of information transferredfrom one signal x to
another signal y and reads:
t(x|y) =∑
p(xi+h,yi,xi) · logp(xi+h|yi,xi)p(xi+h|xi)
(2)
The causality is obtained by comparing the amount ofinformation
transferred from signal x to signal y with theamount information
transferred from signal y to signal x.If tx→y is positive, more
information is transferred from xto y, and therefore x caused y. If
tx→y is negative, y causedx. If it has a value close to 0, then the
causality cannot bededuced.
tx→y = t(y|x)− t(x|y) (3)
This method may give rise to spurious results (Baueret al.,
2007). In order to reduce their number, the ob-tained results are
validated against the plant connectiv-ity information contained in
the topology model of theplant (Thambirajah et al., 2009). CAEX
(Computer AidedEngineering eXchange) is a vendor independent
object-oriented data exchange format based on the XML schemathat
has been used in the past years for structuring planttopology
models in a flexible and expandable way (see e.g.:Schleburg et al.,
2013).
The topology based validation of the data-driven analysisstarts
with the identification of the assets correspondingto the alarms in
the fault template. The connectivity of
the plant is explored using a graph-search algorithm ona graph
capturing the connections between assets in theplant. The depth
first search algorithm is employed inthis work (Yang et al., 2010).
The assets suggested asroot-causes from the data-driven analysis
are taken asstarting points of the disturbance propagation path
whilethe remaining assets are taken as secondary disturbedelements.
All feasible propagation paths between the sug-gested root-cause
and each of the secondary disturbedassets are explored. If no
feasible path is found, the root-cause hypothesis is discarded.
Once the spurious results of the data-driven analysis havebeen
filtered out and one of the root-cause suggestions hasbeen
validated with plant connectivity, the causal alarm istaken as the
alarm associated to the root-cause asset (seeAlg. 3).
Alg. 3: Isolate the causal alarminput : Clusters {Ck} and fault
templates {Tk}output: A causal alarm akc for each fautl template
Tk
1 For each cluster Ck, choose a sequence Ai : Ai ∈ Ck2 {sip} =
Mape→si({ep : ep ∈ Tk})3 Perform process-data causal analysis on
signals {sip} to obtain
suggested root-cause signals {siksg} : q ∈ N4 {asksg} =
Mapsi→as({si
ksg,q}) and {as
ksd} = Mape→as({ep})
5 If ( ∃asksg,q with feasible propagation paths to all asksd
)
akc = Mapas→e(asksg,q)
3. INDUSTRIAL CASE STUDY
The proposed method determining the causal alarm inalarm flood
sequence is illustrated by an analysis con-ducted on the cracking
and quenching section of an ethy-lene plant located in Austria.
The alarm management system of the plant has more than3800
alarms configured in the monitoring system. Theanalysis is
performed using eight months of collected alarmdata, during which
more than 48000 alarms occurred. Theprocess data is presented in an
excel file with a one-second
IFAC DYCOPS-CAB, 2016June 6-8, 2016. NTNU, Trondheim, Norway
726
-
sampling period data set. A CAEX file containing thetopology
model of the plant is used. The P&ID of the plantsection is
shown in Fig. 3. The data are analyzed usinga developed prototype
software tool that guides the userthrough the steps of the proposed
method. The parametersfor the alarm analysis are set by the user.
The thresholdnumber of alarms (τ) (see Section 2.2) is set to a
valueof six. Since operators of the ethylene plant control morethan
one area, the threshold number of alarms is set toa value smaller
than 10 alarms per 10 min. Additionally,the value of 0.4 is chosen
for the clustering threshold aswell as for the prefiltering
threshold (see Subsection 2.3).The parameter σ2 from the MSW
algorithm (Cheng et al.,2013) is set to 5000, allowing alarms with
a time delay ofaround 3.5min to be matched.
Fig. 4. Clustering results: (a) Clusters (b) Abnormalitytemplate
of Cluster1
After loading the alarm log to the software tool and per-forming
the steps Remove chattering alarms and Identifyalarm flood periods,
23 alarm flood periods are identified.Results obtained from the
Cluster alarm flood sequencesstep are depicted in Fig. 4 (a). Three
clusters are obtained.From the 23 alarm sequences analyzed, four do
not belongto any cluster and 19 are clustered under one of the
threegroups. Each one of these clusters containing similar
alarmflood sequences represents a different process
abnormality.Only the process abnormality from Cluster1 is
furtheranalyzed. The template of this abnormality is displayedin
Fig. 4 (b).
To perform the data-driven and connectivity analysis, thesignals
and assets associated to the alarms included in thetemplate are
identified. After loading the process data andthe topology model of
the plant, the automatic mappingusing the Smith-Waterman algorithm
is performed. Theresults are displayed in Fig. 5. Tags in red
represent alarms,tags in blue signals and tags in green assets. For
instance,the valve G02 has an associated signal G02.Y and twoalarms
G02/010 and G02/011.
The process signals appearing in the mapping tree of Fig. 5are
considered for the analysis. These signals are displayedin Fig. 6.
The results of the process-data analysis areshown in a bubble chart
representing the transfer entropybased causality matrix (Bauer et
al., 2007). The analysis
Fig. 5. Alarms (red) from the abnormality template ofCluster 1
mapped to their corresponding signals(blue) and assets (green)
Fig. 6. Time trends of the signals included in the analysis
Fig. 7. Transfer entropy values tx→y
suggests that the root-cause is the signal PC11.XAY.GI01.Y,
PC12.XAY, PI15.Y, PI14.Y, PI13.Y, PI16.Y andGI02.Y are considered
secondary disturbed signals (seeFig. 7).
Finally, the result of the process-data analysis is
validatedusing the process connectivity. The process
connectivityanalysis found feasible propagation paths starting
from
IFAC DYCOPS-CAB, 2016June 6-8, 2016. NTNU, Trondheim, Norway
727
-
Fig. 8. Propagation paths found by the connectivity
anal-ysis
the suggested root-cause to all secondary disturbed points.Fig.
8 depicts the obtained propagation paths. The processconnectivity
analysis suggests that the abnormality spreadfrom the pressure
controller PC11 S to the valves G01and G02 via the information
connection between thiscontroller and the valves. From valve G01,
the disturbancespread to the pressure controller PC12 S.
Simultaneously,the disturbance reached Cracking Zone A, Cracking
ZoneB, Cracking Zone C and Cracking Zone D via the heatexchangers
LQE 1, LQE 2, LQE 3 and LQE 4. SensorsPI13 S, PI14 S, PI15 S and
PI16 S were also affectedby the propagating disturbance since they
are locatedin the coils of the different cracking zones. The
root-cause suggestion from the data-driven analysis is
thereforevalidated by the plant connectivity analysis. An
alarmassociated to PC11 S is the given causal alarm suggestion,i.e.
alarm PC11 CO 4. All the alarms in the template willbe grouped
under it.
After discussion with site engineers the obtained resultswere
confirmed. The analysed abnormal situation corre-sponds to a shift
in the operating mode between cracking-decoking. The transition
between these two operatingmodes triggers a characteristic alarm
flood sequence. Thesite crew mentioned that pressure controller
PC11 is per-forming the opening and closing of both valve G01
andvalve G02 during the change of operation mode and istherefore
the origin of the alarm flood.
4. CONCLUSIONS
This paper proposed an innovative way to reduce alarmflooding by
using a combination of data from differentinformation sources.
Combining alarm log, historical pro-cess data and process
connectivity analysis allowed to notonly group consequence alarms,
but also to determinethe causal alarm of an alarm flood sequence.
The pro-posed method was tested on actual data collected froman
ethylene plant. The proposed approach was able togroup related
alarms triggered during an alarm flood andto identify their causal
alarm.
REFERENCES
ANSI/ISA-18.2 (2009). Management of alarm systems forthe process
industries.
Bauer, M., Thornhill, N., Cox, J., Caveness, M., andDowns, J.
(2007). Finding the direction of disturbance
propagation in a chemical process using transfer entropy.IEEE
Transactions on Control Systems Technology.
Beebe, D., Ferrer, S., and Logerot, D. (2007). Alarm floodsand
plant incidents. ProSys.
Bouchair, N., Gayet, P., and Charbonnier, S. (2013).
Faultisolation by comparison of event lists using a
weighteddistance. IEEE Internation Conference on Systems,Man and
Cybernetics.
Bransby, M.L. and Jenkinson, J. (1998). The managementof alarm
systems: a review of best practice in theprocurement, design and
management of alarm systemsin the chemical and power industries.
HSE ResearchReport CRR 166.
Cheng, Y., Chen, T., and Izadi, I. (2013). Patternmatching of
alarm flood sequences by a modified smith-waterman algorithm.
Chemical Engineering Researchand Design.
Fan, Y. and Deyun, X. (2012). Progress in root causeand fault
propagation analysis of large-scale industrialprocesses. Journal of
Control Science & Engineering.
Henningsen, A. and Kemmerer, J.P. (1995). Intelligentalarm
handling in cement plants. IEEE Industry Appli-cations
Magazine.
Higuchi, F., Yamamoto, I., Takai, T., Noda, M., andNishitani, H.
(2009). Use of event correlation analysisto reduce number of
alarms. Computer Aided ChemicalEngineering.
Hollender, M. and Beuthel, C. (2007). Intelligent alarming.ABB
Review 1/2007.
Hollifield, B. and Habibi, E. (2011). Alarm Management:A
Comprehensive Guide, Second Edition. Isa.
IEC-62682 (2014). Management of alarm systems for theprocess
industries.
Kabir, A., Iyadi, I., Chen, T., Joe, D., and T., B.
(2013).Similarity analysis of industrial alarm flood data.
IEEETransactions on automation science and engineering.
Lai, S. and Chen, T. (2015). Methodology and applicationof
pattern mining in multiple alarm flood
sequences.IFAC-PapersOnLine.
Schleburg, M., Christiansen, L., Fay, A., and Thornhill,N.
(2013). A combined analysis of plant connectivityand alarm logs to
reduce the number of alerts in anautomation system. Journal of
Process Control.
Smith, T. and Waterman, M. (1981). Identification ofcommon
molecular subsequences. Journal of MolecularBiology.
Thambirajah, J., Thornhill, N., Benabbas, L., and Bauer,M.
(2009). Cause-and-effect analysis in chemical pro-cesses utilizing
xml, plant connectivity and quantitativeprocess history. Computers
and Chemical Engineering.
Wang, J., Yang, F., Chen, T., and Shah, S. (2015). Anoverview of
industrial alarm systems: Main causes foralarm overloading,
research status, and open problems.Automation Science and
Engineering, IEEE Transac-tions on.
Yang, F., Shah, S., Xiao, D., and Chen, T. (2012). Im-proved
correlation analysis and visualization of indus-trial alarm data.
ISA Transactions.
Yang, F., Shah, S.L., and Xiao, D. (2010). Signed directedgraph
modeling of industrial processes and their valida-tion by
data-based methods. Control and Fault-TolerantSystems (SysTol).
IFAC DYCOPS-CAB, 2016June 6-8, 2016. NTNU, Trondheim, Norway
728
