Aa

Ma

b

c

a

ARRA

1

pl(ttsm

mmp(saeohu

0h

Nuclear Engineering and Design 249 (2012) 413– 418

Contents lists available at SciVerse ScienceDirect

Nuclear Engineering and Design

j ourna l ho me page: www.elsev ier .com/ locate /nucengdes

decision support system for identifying abnormal operating procedures in nuclear power plant

in-Han Hsieha, Sheue-Ling Hwanga,∗, Kang-Hong Liua, Sheau-Farn Max Liangb, Chang-Fu Chuangc

Institute of Industrial Engineering and Engineering Management, National Tsing Hua University, Hsinchu 300, TaiwanInstitute of Industrial Engineering and Management, National Taipei University of Technology, Taipei 106, TaiwanAtomic Energy Council, Taipei 234, Taiwan, ROC

r t i c l e i n f o

rticle history:eceived 1 June 2011eceived in revised form 8 April 2012ccepted 12 April 2012

a b s t r a c t

In order to prevent safety hazards that can result from inappropriate decisions made by the operators of anuclear power plant (NPP), this study was undertaken to develop a decision support system to reduce thecomplexity of the decision-making process by aiding operators’ cognitive activities, integrating unusualsymptoms, and identifying the most suitable abnormal operating procedure (AOP) for operators. Thestudy was conducted from the perspective of human factors engineering in order to compare the processthat operators originally used to select an AOP with a process that included a support system for AOPidentification. The results of the study indicated that the existence of a support system reduces errors by

quickly suggesting likely AOPs. With such a support system in place, there were clear improvements inhuman performance, i.e., decision-making time decreased by about 25%, and the accuracy of the operators’decisions, judged by the successful resolution of specific problems, increased by about 18%. In addition,there were fewer erroneous solutions implemented, and the mental workload was reduced. Hence, thedecision support system is proposed as a training tool in identifying AOPs in the main control room (MCR).

. Introduction

In nuclear power plants (NPPs), operational safety is the topriority because the release of radioactive materials can result in

oss of life, environmental pollution, and financial losses. Isaac et al.2002) indicated that human error is a major contributor (70–90%)o accidents at NPPs. For instance, the Chernobyl disaster in 1986 inhe Soviet Union was caused by human errors that dealt with deci-ion making, receiving information, and action selection. In order toaintain safety in NPPs, the human factor issue cannot be ignored.Alarm systems play an important role in NPPs. They are the

ain source that operators must depend on for detecting abnor-al situations and failures. Alarm systems monitor all important

lant systems and alert operators when abnormal situations occurO’Hara et al., 2000; Noyes and Bransby, 2001), so robust alarmystems are very important. In conventional NPPs, the traditionalnalog alarm systems use continuous physical phenomena, such aslectrical, mechanical, or hydraulic quantities, to assess the status

f the operational components of the plant. However, as technologyas developed, digital systems gradually have been overcoming these of analog systems in NPPs. Although digital systems are more

∗ Corresponding author. Tel.: +886 35742694; fax: +886 35722685.E-mail address: slhwang@ie.nthu.edu.tw (S.-L. Hwang).

029-5493/$ – see front matter © 2012 Elsevier B.V. All rights reserved.ttp://dx.doi.org/10.1016/j.nucengdes.2012.04.009

© 2012 Elsevier B.V. All rights reserved.

precise, too much information and too many alarms may make itdifficult for the operator to comprehend what is actually happeningin the plant. For this reason, some latent problems associated withthe digital main control systems must be solved.

In an NPP, the operators’ tasks include information gathering,planning, decision making, and avoiding unforeseen risks throughthe alarm system (Noyes and Bransby, 2001) if they follow an incor-rect operating procedure, for instance. The operators’ tasks are tomonitor the system continually to ensure that the system is stableand functioning normally (Ma et al., 2006). During abnormal situ-ations, a well-trained operator should comprehend a malfunctionin real time by analyzing alarms, assessing values, or recognizingunusual trends of multiple instruments (Hogg et al., 1995; Vicenteet al., 1996). In an NPP, many alarms from many different systemsoften occur at the same time during an incident, making it difficultfor the operator to select a correct response efficiently. Too manyinformation imposes a heavy burden on operators in a time-criticalsituation, and it is very difficult for them to conduct a thoroughassessment of each individual symptom in a short period of time.In the absence of operator support systems, the operators must con-sider an overwhelming amount of information and make decisions

very quickly. Unfortunately, this can take too much time. Since thedecision-making environment is extremely complicated and dataintensive, the use of automated systems or expert systems to aiddecision making is likely to become more common. The models

4 eering

ose1

ricAmOaafiblEaaastesbssocoatAdpmmtArateha

2

aivowticpeKafipod

14 M.-H. Hsieh et al. / Nuclear Engin

f these cognitive agents must be accurate and complete if suchystems are to be effective in enhancing the speed, accuracy, andfficiency of human decision making (Parasuraman and Mouloua,996).

The Lungman NPP in Taiwan uses an advanced boiling watereactor (ABWR) and digital alarm systems. Hence, in this study, thessues at the Lungman NPP were used as a case in order to dis-uss the design and safety of the fault diagnosis and alarm systems.t the Lungman NPP, three kinds of procedures have been imple-ented to deal with an abnormal operation status, i.e., Emergencyperating Procedure (EOP), Abnormal Operating Procedure (AOP),nd Annunciator Response Procedure (ARP). There are seven EOPs,bout 80 AOPs, and over 20,000 ARPs. When the alarms occur, therst thing the operator must do is to confirm whether an EOP shoulde performed and, if so, which one. If the plant is at risk for radiation

eakage, the operator will perform an EOP. If there is no appropriateOP for the emergency situation at hand, the operator will considern AOP next. If there is no proper AOP, the operator will operatelarms with ARPs. The entry conditions of EOPs are clearly definednd listed in the EOPs. In addition, each individual alarm has apecific ARP. However, identifying AOPs is more difficult than iden-ifying EOPs or ARPs. There are many abnormal symptoms listed inach AOP. All of the symptoms listed in the AOP might not neces-arily be presented. Some symptoms may be related to each other,ut some are independent of each other or mutually exclusive. Theituation becomes more difficult when many AOPs have the sameymptoms. The operators must consider many information sources,rganize information, and make a decision. Wickens (1992) indi-ated that there are four limitations that influence the quality of theperator’s decision, i.e., perception, attention, long-term memory,nd working memory. It is evident from the above description thathe process of decision making that involves the identification ofOPs is very complex. For this reason, this study is focused on theevelopment of a decision support system to improve operators’erformance and to lower the chance of human errors. There areany formal and informal symptoms that help plant operators toake their decisions. In the present study, we focus on the symp-

oms provided by the alarm system for selecting the most likelyOP in time. The objectives of the decision support system are toeduce the amount of information that an operator must considernd integrate during the abnormal situation, to ensure that opera-ors do not overlook important information, and to reduce humanrrors. In addition, the study was conducted from the perspective ofuman factors engineering in order to validate the appropriatenessnd effectiveness of the system.

. Decision support system applications in NPPs

Kim et al. (2001) used an alarm and diagnosis-integrated oper-tor support (ADIOS) system to prevent too many alarms fromnfluencing the operator’s judgment. They indicated that the acti-ation of a large number of alarms imposes too heavy a burden onperators in such a time-critical situation. Multiple alarms interfereith the operator’s judgment and ability to diagnose the situa-

ion, and they are likely to contribute to human errors. Artificialntelligence techniques have the potential to make a significantontribution to the reliable operation of NPPs, and there were manyrevious studies concerning the design and implementation ofxcellent expert systems and operator support systems. Kwon andim (1999) applied the Hidden Markov Model (HMM) to identifyccidents in NPPs and showed its robustness. The accident identi-

cation system accurately identifies the type of accident and alsoredicts abnormal occurrences in advance. Lee et al. (2007) devel-ped the fault diagnosis advisory system (FDAS), which is based onynamic neural networks. They indicated that FDAS facilitates the

and Design 249 (2012) 413– 418

fault diagnosis task and reduces errors by quickly suggesting appro-priate courses of action. FDAS provides accurate, reliable advice foroperators who must make decisions quickly.

However, according to several papers that assessed the evalua-tion results provided by decision support systems, such systemsdo not guarantee improvement in the operator’s performance.Some support systems could actually increase the operator’s men-tal workload during these critical times. Therefore, an effectivesupport system must aid, not hinder, the operator’s cognitive pro-cesses (Yoshikawa, 2005; Kim and Seong, 2006). Time constraintsand the volume of information also are crucial issues that sup-port systems must deal with. For any initiating event, there willbe a time period beyond which a bad consequence may result. Inthat time period, the operator has the opportunity to take someaction to prevent an adverse radiological event or a reactor acci-dent. The time period might range from fractions of a second to afew minutes. However, if the allowable time period is too short,it is very difficult for the operator to make a decision. In addition,the quantity of information available may impose a heavy burdenon operators in a time-critical situation. Thus, a decision supportsystem should be capable of repeating and updating informationcontinually and it should be capable of processing multiple prob-lems simultaneously. Most importantly, the support system shouldbe programmed to provide the best possible answer based on thesituation at hand. The combination of a human decision maker andan automated decision aid should result in improved performance(Bernard, 1999).

3. Method and experiment

After visiting the operation training in the Lungman nuclearpower plant for several times and interviewing with operators,operators proposed the difficulty of selecting an appropriate AOPunder time pressure. Therefore, this study focused on constructinga decision support system to resolve the current problem.

3.1. Decision support system

A decision support system could filter out unrelated informa-tion and AOPs and gather some symptoms related to the currentevent. In addition, it could provide the operator with some suit-able AOPs. The operator could refer to the advice of the supportsystem and consider what to do next. According to the AOPs, thereare many symptoms related to abnormal events, but it is not neces-sary to present every symptom listed in the AOPs. Moreover, sincesome symptoms might belong to several AOPs, the support sys-tem can select the related AOPs when an unusual symptom occurs.After the similarity analysis, the operator can refer to the infor-mation for decision making. Because all the decision rules of thesupport system are based on abnormal operating procedures, themore complete AOPs make the decision support system more pow-erful.

3.1.1. Constructing the abnormal symptom databaseBefore constructing a decision support system, a database that

contains the symptoms of abnormal events must be constructed.The steps for constructing such a database are as follows: (1) col-lecting all the abnormal symptoms from the AOPs; (2) classifyingthe symptoms by information source type; and (3) constructingthe abnormal symptom matrix. In step 3, the abnormal symptommatrix can be constructed by using the attributes of the informa-tion. It is straightforward to place the information sources of system

level alarms, plant level alarms, plant status tiles, and individualalarms into “symptom occurs” and “symptom does not occur” cat-egories. In the abnormal symptom matrix, the “symptom occurs”category is designated by “1”, and the “symptom does not occur”

eering

cps

X

tot1n

3

isFttp

˚

nsrfiAstrlsIt

�

ostAlIA

3

(twbw

M.-H. Hsieh et al. / Nuclear Engin

ategory is designated by “0”. After transformation, the data can berocessed by the computer program. The structure of the matrix ishown below:

(m×n) =

⎡⎣

x11 · · · x1n

.... . .

...xm1 · · · xmn

⎤⎦ (1)

X(m×n) is an abnormal symptom matrix. The symbol xij refers tohe value of symptom j in AOPi. The “m” is the number of abnormalperating procedures, and “n” is the number of abnormal symp-oms. The i and j identify the index. All the values of xij are either

or 0. If xij = 1, symptom j exists in AOPi. If xij = 0, symptom j doesot exist in AOPi (Table 1).

.1.2. Decision making processThere are two stages in the decision making process, i.e., filter-

ng and similarity analysis. The objective of the filtering stage is toelect all the AOPs that are related to the current status of the plant.irst, the current abnormal symptoms of the plant are imported tohe support system. After comparing the status of the plant withhe abnormal symptom matrix and deleting unrelated AOPs, theossible AOPs are identified.

(p×q) =

⎡⎢⎢⎢⎣

z11 z12 · · · z1q

z21 z22 · · · z2q

......

. . ....

zp1 zp2 · · · zpq

⎤⎥⎥⎥⎦ (2)

The above matrix �(p×q) is the result after filtering. The p is theumber of the symptoms that occur that are related to AOP. Theymbol q is the number of AOPs filtered (q ≤ n, p ≤ m). The term zepresents whether the symptom occurs in the AOPs after they areltered. All the values of zij are 1 or 0. If zij = 1, symptom j exists inOPi. If zij = 0, symptom j does not exist in AOPi. After the first stage,ome AOPs may be selected. The other stage is similarity analysis,he objective of which is to find the most suitable AOP from theesults of the filtering stage. There are three indices in the simi-arity analysis, i.e., the number of symptoms, the influence of theymptoms, and the correlations that exist between the symptoms.n this study, the correlation between the symptoms was assumedo be independent.

=

⎡⎢⎢⎢⎢⎢⎣

∑qj=1wjz1j∑qj=2wjz2j

...∑qj=pwjzpj

⎤⎥⎥⎥⎥⎥⎦

=

⎡⎢⎢⎢⎣

�1

�2

...�p

⎤⎥⎥⎥⎦ (3)

The similarity of each AOP filtered in the second stage can bebtained by combining the three indices. The above matrix is theimilarity matrix �. In the matrix, wj represents the weight of symp-om j. The z represents whether the symptom occurs or not in theOPs filtered in or selected. The �i is the similarity of AOPi. The

arger the value of �i, the higher the similarity of AOPi becomes.t is suggested that the operator consider the highest similarity ofOP first.

.1.3. Output of the decision support systemThe output of the decision support system was displayed on VDU

Video Display Units). The following is the information provided to

he operator by the decision support system: (1) all the AOPs thatere selected by the decision support system; (2) the relationship

etween the current plant status and the AOPs that were selected,hich is helpful to the operator in the decision making process; and

and Design 249 (2012) 413– 418 415

(3) the similarity of the AOPs that were selected by support system,which is helpful to the operator for deciding the most suitable AOP.

3.2. Experiment

The purpose of the experiment was to evaluate the performanceand effect between original mode and the support system mode.In addition, the question concerning whether the decision supportsystem can significantly lower the operator’s mental workload andimprove performance was answered in the affirmative.

3.2.1. ParticipantsThe participants were 32 graduate students from College of

Engineering at National Tsing Hua University. Their backgroundsare similar to that of new operators in NPPs. The participants rangedin age from 22 to 29 years old. All of the participants had normaleyesight and were in good health. In order to make their basicknowledge similar to that of the NPP operators, the participantswere asked to study basic operation information related to NPPs,memorize some important information, and take a qualificationtest before the experiment. In the qualification test, the subjectswere asked to classify the system level alarms to reactor-related,cooling system-related or power system-related. The informationwas intended to help the subjects realize the situation and makedecisions in the experiment. The formal experiment did not startuntil the subjects were able to classify up to 90% system level alarmin the qualification test.

3.2.2. Experimental designAll the subjects had to use two different operating modes. The

order of the operating modes was distributed randomly. There weremany decision making tasks in each mode of the experiment, andthe subjects had to determine the cause of the unusual state andidentify the abnormal operating procedure. The experiment lastedabout 50 min, and the data of decision time, error ratio, and NASA-TLX score were collected.

There were two independent variables in the experiment, i.e.,operating mode and the number of abnormal events. Each inde-pendent variable included two levels, i.e. (1) operating mode, theoriginal mode, and the support system mode and (2) the number ofabnormal events happened, i.e., one or two. After consulting withprofessional operators, thirteen abnormal events were simulatedin advance. The number of common symptoms between any twoabnormal events was from 0 to 2. The subjects did not know howmany events might occur during the experiment.

For the dependent variable of decision time, the number oferrors and the NASA-TLX Index for cognitive workload were mea-sured in the experiment. The definition of decision time was theduration that the scenario started until subject made a decision.The number of errors is the wrong decision subject made. NASA-TLX Index was obtained by the NASA-TLX questionnaire including 6dimensions, mental demand, physical demand, temporal demand,effort, performance, and frustration level.

3.2.3. Experimental environmentThe experiment was carried out in the laboratory. The WDP was

simulated by three projection screen (Fig. 1). For the experimentalenvironment and the layout of the original mode, a system-levelalarm display, a mimic-parameter display, and a book of abnormaloperating procedures were provided. In the support system mode,the advice of the decision support system was displayed on a VDU.

The information on the top of the screen was the unusual symptomscurrently and the AOPs which were selected. The information ofAOPs’ assign number and the related information of mimic param-eter to AOPs were in the middle of screen. All the AOP filtered in

416 M.-H. Hsieh et al. / Nuclear Engineering and Design 249 (2012) 413– 418

Table 1Partial of the abnormal symptom matrix.

RPS RBSW RBCW SGT LDI ARM PRM CSTF OG RW FPCU ACS DWC SLC RCIC HPCF RHR RBHV TBHV . . .

501.1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 . . .501.2 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 . . .501.3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 . . .501.4 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 . . .502.1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 . . .502.2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 . . .503.1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 . . .503.2 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 . . .504.1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 0 0 . . .505.1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 . . .506.1 0 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 . . .507.1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 . . .508.1 0 0 0 0 1 1 0 0 0 0 0 1 1 0 0 0 0 0 0 . . .508.2 0 0 1 0 1 0 0 0 0 1 1 0 0 0 0 0 1 1 0 . . .508.3 0 0 0 0 1 0 0 1 1 1 0 0 0 0 0 0 0 0 1 . . .508.4 0 0 0 0 1 1 1 0 0 0 0 0 0 0 0 0 0 1 0 . . .508.5 0 0 0 0 1 1 1 0 0 1 0 0 0 0 1 0 0 1 0 . . .508.6 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 . . .509.1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 . . .509.2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 . . .509.3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 . . .509.4 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 . . .510.1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 . . .511.2 0 0 0 1 1 1 1 0 0 0 0 0 0 0 0 0 0 1 0 . . .512.1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 . . .512.2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 . . .512.3 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 . . .

.

.

....

.

.

....

.

.

....

.

.

....

.

.

....

.

.

....

.

.

....

.

.

....

.

.

....

.

.

....

. . .

ed by

wt

4

Anatwefwwn

4

astt

significantly different (Z = −3.740, p < 0.05). Fig. 2 indicates that theaverage decision-making time of the support system mode wasabout 77.6% of the average decision-making time of the originalmode when only one event occurred. When two events occurred,

Fig. 1. The WDP simulat

ere ranked by similarity, and the most suitable AOP was listed onhe top, and so on.

. Results

In order to confirm that the data were normally distributed, thenderson–Darling test was used to check whether the data fit aormal distribution (p > 0.05) or not. The results of the normal prob-bility tests indicated that only one of three dependent variables fithe normal distribution (p > 0.05). The three dependent variablesere: (1) the decision-making time (p < 0.005); (2) the number of

rrors (p < 0.005); and (3) the NASA-TLX scores (p > 0.05). There-ore, the hypothesis of normality for NASA TLX scores was accepted,hereas those for decision-making time and the number of errorsere rejected. Hence, the data of decision-making time and theumber of errors were analyzed with non-parametric tests.

.1. Decision-making time

The decision-making time was the duration between when the

larm signal started and the subject completed her or his deci-ion. The results of the Wilcoxon signed-rank tests indicated thathe median of the decision-making time of the original mode andhe support system mode under one abnormal event occurrence

three projection screen.

were significantly different (Z = −2.206, p < 0.05). In addition, themedian of the decision-making time of the original mode and thesupport system mode when two abnormal events occurred were

Fig. 2. Comparison of the decision-making time of the different modes.

M.-H. Hsieh et al. / Nuclear Engineering and Design 249 (2012) 413– 418 417

F

t7

4

matr1Wrstm(msofaw

4

ajsitHse

4

otiTs(

mal event is simpler to deal with, and the subjects were less likelyto be confused. There were no significant difference in the numberof mistakes between the two modes, but there were fewer misserrors in the support system mode. Moreover, the support system

Table 2Comparison of the two modes.

Comparison item Mode type

Original mode Supportsystem mode

Accuracy Total 65.5% 83.3%One event No significant differenceTwo events 51.6% 75.0%

Error type Miss More LessMistake No significant difference

Decision-making time One event Long Short

ig. 3. Comparison of the accuracy of decision making for the different modes.

he average decision-making time of the support system mode was5% of the average decision-making time of the original mode.

.2. The accuracy of decision making

All the subjects were assigned to the two different operatingodes in random order. Each group conducted two kinds of situ-

tions with different numbers of abnormal events. Fig. 3 indicateshe accuracy of their decisions. When one event occurred, the accu-acy of decision making was 93.8% with the original mode and00% with the support system in the experiment. The result of theilcoxon signed-rank tests confirmed that the median of the error

atio of the original mode was not significantly different from theupport system mode (Z = −1.414, p = 0.157). On the contrary, whenwo events occurred, the median of the error ratio of the original

ode was significantly different from the support system modeZ = −2.588, p < 0.05). In the experiment, the accuracy of decision-

aking was about 51.6% with the original mode and 75% with theupport system. When both situations were combined, the medianf the overall error ratio of the original mode was significantly dif-erent from the support system mode (Z = −2.911, p < 0.05). Theccuracy of the decision-making support system mode was 65.5%ith the original mode and 83.8% with the support system.

.3. The error type of decision-making

There were two types of errors, i.e., mistakes and miss. Mistakesre errors that indicate incorrect planning of actions, i.e., the sub-ects selected the wrong procedure. A “miss” error refers to theituation in which the subjects did not select any procedure whent was necessary to choose one. The number of mistakes betweenhe two modes was not significantly different (Z = −1.311, p > 0.05).owever, the number of miss errors between the two modes was

ignificantly different (Z = −2.556, p < 0.05). There were less missrrors in the support system mode.

.4. Subject workload rating NASA-TLX

A paired-sample t-test was conducted on the experimental dataf subjective workload rating (NASA TLX scores). The result showedhat the median of differences between the two different operat-

ng modes was statistically significant (t = −3.02, p < 0.05). NASALX scores for operating in the original mode (M = 60.84) wereignificantly higher than operating in the support system modeM = 50.01), as shown in Fig. 4.

Fig. 4. Comparison of NASA-TLX scores for the two modes.

In the experiment, about 6% of the subjects said that there wasno significant difference between the two modes. About 72% of sub-jects preferred operating with the fault diagnosis support system,and they thought the support system mode was helpful for decisionmaking. However, only 22% of subjects preferred operating in theoriginal mode.

5. Discussion

5.1. The comparison of operating modes

Wickens (1992) indicated that there are four limitations thatinfluence the quality of diagnosis, i.e., perception, attention, long-term memory, and working memory. The decision support systemhas been designed to integrate the information and list the pos-sible AOPs so that the mental workload of the operators couldbe reduced. That is the reason the statistical tests indicated thatthe support system mode was significantly better than the originalmode in almost all aspects.

Table 2 shows the comparison of the two modes. The resultsshow that the accuracy of decision-making with the decision sup-port system was higher than that with the original mode. Subjectswith the support system spent less time making a decision thanthose in the original mode, irrespective of whether one or twoabnormal events occurred. Especially, when two abnormal eventsoccurred, operating with the support system reduced the decision-making time and also improved the accuracy of the diagnosis.However, there was no significant difference in the accuracy of thedecision making between the two modes when only one abnormalevent occurred. The reason might be due to the fact that one abnor-

Two events Long Short

Mental workload High LowSubjective preference Minority Majority

4 eering

wdiastsaaat

missirsftmd

5

5

entAwTmFes

5

tarseeAno

6

siea

ts

18 M.-H. Hsieh et al. / Nuclear Engin

as able to help people integrate all the information during theecision making process. That was why the subjects were able to

dentify crucial abnormal symptoms easier. When the number ofbnormal events increased, the operators made more errors irre-pective of the operating mode. Concerning the mental workload,he subjective rating on NASA-TLX in the support system mode wasignificantly lower than that in the original mode. It is difficult forn operator to make a correct decision when he or she is working in

situation that imposes a high mental workload. That is why almostll the subjects’ performances with the support system were betterhan those with the original mode.

Conrad (1951) indicated that too much information leads toore time-consuming filtering at the expense of decision qual-

ty. Reducing the load on working memory may make a correctelection easier. Kim and Seong (2006) indicated that, if an unusualituation imposes a high memory load, an operator may choose annformation-economic strategy. However, if a support system canelieve the memory load, the operator can shift to a more reliabletrategy. That may be the reason that almost all the subjects pre-erred the support system. Moreover, the support system helpedhe subjects obtain the correct procedure efficiently, and, as a result,

ost of the subjects thought the support system was useful forecision making.

.2. Study limitations

.2.1. Background of subjectsBecause it was not easy to invite operator to participate the

xperiment, the participants were all graduate students who hado experience of complex alarm system in the control room. Dueo the subjects’ capabilities limitation, there were only thirteenOPs out of eighty-five AOPs in the experiment. All the subjectsere trained and practiced to be familiar with all AOPs and alarms.

herefore, all subjects could be treated as potential operators toimic the diagnostic operating in the simulation control room.

or future study, it needs to design experiments with experi-nced operators, who are able to handle more alarms in abnormalituations.

.2.2. The database of the support systemThere are over eighty AOPs, and each procedure includes six

ypes of abnormal symptom. Due to the large amount of data, thebnormal symptom database only contained system level alarmselated to all the AOPs in the study. All the correlations betweenymptoms were assumed to be independent and the weight ofvery symptom was assumed to be the same in this study. How-ver, the prototype of support system still can filter out all possibleOPs, and advise the most suitable AOPs to subjects. The complete-ess of the abnormal symptom database is concerned to the robustf the support system.

. Conclusions

The goals of this study were as follows: (1) to construct a deci-ion support system for the purpose of quickly and accuratelydentifying abnormal operating procedures and preventing humanrrors and (2) to verify the decision support system’s effectiveness

nd its ability to reduce mental workload.

The results of experiment verified that the subjects’ decisionimes were reduced significantly when they were assisted by theupport system. Almost all participants took less time in identifying

and Design 249 (2012) 413– 418

the AOPs when an abnormal event occurred, which left additionaltime for the operator to deal with abnormal event. In addition, theerrors of procedure selection and subjects’ mental workload canbe reduced with the assistance of the support system. The mostimportant observation is that the support system can help peopleavoid overlooking important information when several abnormalevents occur simultaneously. In conclusion, we highly recommendthe decision support system because it is helpful in identifying theAOPs for system level alarms.

In order to enhance operating safety, future research shouldfocus on three different parts of the decision support system, i.e.(1) A field study with real operators using this decision supportsystem is necessary to confirm the feasibility of the support sys-tem, (2) Teamwork with the support system. More studies may berequired to confirm the effect of the decision support system whenoperators work as a team to identify abnormal operating proce-dures, and (3) the robustness of the support system. The weights ofevery symptom are important to reach the most accurate advice.By the nuclear engineering expert’s assessment and previous oper-ating records analysis, the suitable weights of every symptom maybe estimated to improve the support system in the future.

Acknowledgments

This study was supported by the ROC Atomic Energy Council andthe ROC National Science Council, project no. NSC99NUE007008.The authors would like to thank the administration of LungmanPower Plant in Taiwan providing valuable information.

References

Bernard, J., 1999. Issues regarding the design and acceptance of intelligent supportsystems for reactor operators. IEEE Trans. Nucl. Sci. 39 (5), 1549–1558.

Conrad, R., 1951. Speed and load stress in a sensori-motor skill. Br. J. Ind. Med. 8,1–7.

Hogg, D.N., Folleso, K., Volden, F.S., Torralba, B., 1995. Development of a situationawareness measure to evaluate advanced alarm systems in nuclear power plantcontrol rooms. Ergonomics 38 (11), 2394–2413.

Isaac, A., Shorrock, S.T., Kirwan, B., 2002. Human error in European air traffic man-agement: the HERA project. Reliab. Eng. Syst. Safe. 75, 252–272.

Kim, J.H., Seong, P.H., 2006. The effect of information types on diagnostic strategiesin the information aid. Reliab. Eng. Syst. Safe.

Kim, J.T., Kwon, K.C., Hwang, I.K., Lee, D.Y., Park, W.M., Kim, J.S., Lee, S.J., 2001. Devel-opment of advanced I & C in nuclear power plants: ADIOS and ASICS. Nucl. Eng.Des. 207, 105–119.

Kwon, K.C., Kim, J.H., 1999. Accident identification in nuclear power plants usinghidden Markov models. Eng. Appl. Artif. Intell. 12, 491–501.

Lee, S.J., Mo, K., Seong, P.H., 2007. Development of an integrated decision support sys-tem to aid the cognitive activities of operators in main control rooms of nuclearpower plants. In: Proc. of IEEE Symposium on Computational Intelligence inMulticriteria Decision Making (MCDM), pp. 146–152.

Ma, R., Kaber, D.B., Jones, J.M., Starkey, R.L., 2006. Team situation awareness innuclear power plant process control: a literature review, task analysis and futureresearch. In: ANS Fifth International Topical Meeting on Nuclear Plant Instru-mentation, Controls, and Human Machine Interface Technology, Albuquerque,November, pp. 12–16.

Noyes, J.M., Bransby, M., 2001. People in Control: Human Factors in Control RoomDesign. IEE Press, London, pp. 207-221.

O’Hara, J.M., Brown, W.S., Hallbert, B., Skraning, G., Persensky, J.J., Wachtel, J., 2000.The Effects of Alarm Display, Processing and Availability on Crew Performance(NUREG/CR-6691). U.S. Nuclear Regulatory Commission, Washington, DC.

Parasuraman, R., Mouloua, M., 1996. Automation and Human Performance: Theoryand Applications. Lawrence Erlbaum Associates, Mahwah, NJ, pp. 183–197.

Vicente, K.J., Moray, N., Lee, J.D., Rasmussen, J., Jones, B.G., Brock, R., Djemil, T., 1996.Evaluation of a Rankine cycle display for nuclear power plant monitoring and

diagnosis. Hum. Factors 38 (3), 506–521.

Wickens, C.D., 1992. Engineering Psychology and Human Performance. HarperCollins, New York.

Yoshikawa, H., 2005. Human–machine interaction in nuclear power plants. Nucl.Eng. Technol. 79 (2), 151–158.