593

1 INTRODUCTION

Regardless on doubt of the famous statement that said

80% of maritime incident are caused by the human

error, the issue of human contribution to accidents is

still remain important to study in the maritime safety

area [1]. Departing from the traditional approach of

studying human error for finding the lack of human

limitation and blaming them, human factors study

needs to be seen as the area where the human

limitation is studied to improve working conditions.

Thus, paradigm shift is needed to enhance the

understanding of human capabilities, further, to

facilitate the development of countermeasure and

preventive strategy.

For the specific human performance factors that

caused the most incident, situation awareness (SA) is

prominent substituent that often mentioned. Human

error in SA is labeled as “loss of SA”. Incident analysis

report that was conducted by Grech et. al revealed

71% of incidents in maritime operator are due to loss

of SA [2]. Loss of SA will result in the operator failure

to understand the operation condition and leading to

the failure to take appropriate action. For thus, as the

well concept that not only researcher and

management define but also seafarer well understand

and relate its importance, IMO already included it in

crew resources management as one of the non-

technical skill for seafarer to have [3,4].

1.1 Human Reliability Analysis in Maritime Operation

Practice to quantify human error through the process

of human reliability analysis (HRA) has become the

practice in other industry, from nuclear power plant

through the air traffic control [5]. To uniform it, IMO

with the formal safety assessment (FSA) guideline has

also mentioned the HRA and various method of it in

Simulator-based Human Reliability Analysis using

Bayesian Network: A Case Study on Situation

Awareness in Engine Resources Management

A.M. Nizar, T. Miwa & M. Uchida

Kobe University,

Kobe, Japan

ABSTRACT: Situational awareness (SA) is regarded as one of the important non-technical skills in constructing

the seafarers’ ability in daily decision-making and performing tasks, especially in Engine Resources

Management (ERM). In maritime accidents that are mainly human error, insufficient SA is the specific factor

that contributes to most of incidents. To quantitatively assess SA reliability, a Bayesian network of seafarers’

performance in attaining SA in engine supervisory control is constructed. The adaptation of simulator data

helped as combination along using subject matter expert input, which is a common practice in constructing

human reliability analysis. Additionally, the simulator data can serve as the updating function when new data

is observed. The result shows that the model can provide promising results as compared with expert

expectation. Such kind of model can support the evaluation of the engine operation onboard, and mitigation can

be provided to reduce the probability of human error.

http://www.t

ransnav.eu

the

International Journal

on Marine Nav

igation

and Safety of Sea Transpor

tation

Volume 18

Number 3

September 2024

DOI: 10.12716/1001.18.03.1

594

maritime operation [6]. The HRA is not only useful in

academic activity, but it was found to be benefit for

stakeholders such ship owners and safety inspectors

to identify and minimize the potential risk. The

concept of hardware reliability analysis was also

implemented in the HRA methods, including the

hazard analysis and risk control stage. Seeing the

HRA as similar with hardware analysis, it is also

believed similar two combination of method can be fit

to each other and combined to measure system

reliability. Various HRA methods have been

developed, also demonstrated in maritime operation

cases. Cognitive reliability and error analysis

(CREAM) is already applicated to derived the marine

engineers performance reliability by combining

Bayesian inference and fuzzy methods [7]. In seeing

future projection, fuzzy methods also used combine

with success likelihood index methods (SLIM) to

demonstrate the autonomous operation regarding the

human-machine interface [8]. Beyond onboard, the

SLIM methods combined with system theoretic

process analysis (STPA) are also used to analysis

human-machine interaction in ship-to-ship LNG

bunkering [9].

Most HRA aims to measure human error

probability (HEP), that defined as an index to show

the likeliness the human will conduct an error during

the specific event. IMO defines it as the ratio of

number of human error that have occurred, with the

number of opportunities for human error [6]. In

general, denominator for HEP is the number of

chances the human conduct the error, compared to

the hardware reliability where the denominator is the

running time of equipment. This led to quantifying it

by bottom-up approach to predict HEP by retrieving

various data, mostly accident or incident report. The

pitfall of employing only the failure database is the

information only contain the number of failure event,

without number of successful performance, where it is

more close to assess the failure probability with

empiric way [10]. Deciding the human error data from

the accident report also has limitation since the

number of accident reports is considered small

compared to hardware failure data. This limitation

often counters by including the expert judgement as

the input, or the simulator experiment data.

In the lower factors, HRA can be included,

combine, and consist of several performance shaping

factors (PSF)[6]. PSF is often treated independently

from one to the other. Several agree that PSF can be

overlap, or its influence to each other should be

considered [11]. The countermeasure of the

dependency issue between PSF is by utilizing

Bayesian network. The utilization of Bayesian

network in HRA is increasing recently [5]. The

Bayesian network allows us to analyze the likelihood

of human error and identify the dependencies for

complex modelling. It also came with the advantage

of the ability to combine various data.

Bayesian network utilization in maritime operation

is steadily increasing, either for HRA purpose or

system reliability. It has wide application range from

operator safety assessment to the evacuation training,

including its application in offshore operation [12,13],

ship collision [14], emergency situation [15], and ship

engine operation [7]. The Bayesian network suggest

PSF interaction and integrating different sources of

information into the model, once the new information

or data is exist, it can be updated easily to the model

[11]. In the context of HRA, the Bayesian network

provides the ability to contain and combine multiple

types of information and data, including cognitive

literature, operation events, statistical data, and expert

judgment.

SA concept in human factors field is already

applied in various work environments, include in

maritime operation. In this study, loss in SA is

considered as one of the factors that contribute to the

human error event. HRA as the methods in

quantifying human error is applied with adapting the

engine plan simulator data combine with the subject

matter experts. Further, Bayesian concept is employed

to accommodate the dependency between the factors

in contributing the human error.

2 MODEL CONSTRUCTION AND UPDATING

There are two terms involved in model construction

and updating. In model construction, Bayesian

network is applied to construct the model and

calculate the probability of SA failure in the first place.

While in model updating, the concept of Bayesian

inference is used to recalculate the probability of SA

failure by considering the new data from the

simulator. The difference between the two methods is

Bayesian network with its graphical methods does not

necessarily imply the theoretical Bayesian inference.

However, the Bayesian network in this study is called

Bayesian since it employed similar rules for inferring

the probability.

2.1 Model construction

The flow process of model construction and updating

are shown in Figure 1. The Bayesian network in

calculating human error mostly employed subject

matter experts input as weighting factors, in this flow

is to calculate nodes distance. Beyond that, subject

matter experts also use variable control input to

calculate condition probability distribution. Within

this study, the role of subject matter expert is not

removed, but instead reduced by substituting it with

the result from simulator data, specifically in the stage

of calculating the condition probability distribution.

Bayesian network causal model uses directed acyclic

graphs consisting of nodes and arcs. The node plays

as the variable in the model, and the arcs denote the

causal relationship between these variables. The

nodes that the arcs point to are called child nodes,

while the reference nodes are called parents nodes. A

node with child node and no parent node is called

root node. As shown in Figure 1, this part refers to the

first and second stage. For nodes that are discrete,

their effect on the child node can be quantitively

expressed through a conditional probability

distribution (CPD) that shows the influence of parent

nodes. This part takes the three processes on the last

section of model construction.

595

Figure 1. Flow of model construction and updating.

Figure 2. Bayesian network general construction

2.1.1 Identify nodes and causal relationship

The Bayesian network model has several

advantages, such as the ability to clearly define causal

factors. Figure 2 illustrates the causal factor mapping

for the proposed model, which focuses on estimating

the likelihood of human failure event (HFE) as the

general objective of human factor analysis. Based on

the simulator data, this study focuses on the Bayesian

network depicted in Figure 2 in general and Figure 3

as the proposed model, which include three level of

stages: failure mode, cognitive function, and

performance shaping factors.

Figure 3. Bayesian network based on the simulator

experiment

Failure Mode (FM). HFE can be described as many

types as possible of activity or process during the

work. Following the same term in hardware

reliability, FM is constructed below the HFE to

explain various types of supporting events that lead

to the human error itself. This can range from

processing information process, decision making

process and taking an action process. In this study, SA

loss is assigned as one of the FM that construct the

HFE.

Cognitive Function (CF). FM are supported by

various human mechanisms explained by CF nodes.

CF is determined as a variable that can be examined

and observed during the controllable environment

such as simulator, but it is difficult to observe in work

environment. As the simulator experiment design, CF

is assigned as one or multiple dependent variables. As

shown on the Figure 3, there are two CF nodes called

perception and comprehension, that support the SA

loss. Perception, as situation awareness definition, is

the process consuming the cues from the

environment, while the comprehension is the stage

where this information from the cues combined with

the working and long-term memory to create

meaningful information for the work goal.

Performance Shaping Factors (PSF). The PSF are the

lowest, or the root nodes of the modes, as other HRA

methods also described. In relation to the simulator

data, the PSF represents the independent variable, in

this case task-load, expertise, and familiarity. Task-

load refers to the complexity of the performed work,

the expertise is determined by the skill and

knowledge of the operator, and familiarity refers to

experience with specific work environment hardware

and scenario. These PSF nodes contribute to the SA

loss by influencing the operator perception and

comprehension.

2.1.2 Distance between nodes

The relative importance of each parent node in

influencing the child nodes are established by

considering the relative importance of one parent

node compared to others. Røed suggested that this

can be done using a weight w

i for each parent node i

[12]. In this study, we employed analytical hierarchy

process to conduct pairwise comparisons and

determine the weight. We invited four experts from

academia who have professional experience on board

in engine department as described on Table 1. The

experts provided their weighting answers

individually.

Table 1 Profile of the expert as subject matter experts

________________________________________________

No. Professionals Onboard Duration

________________________________________________

1 Academics Chief Engineer 15 years

2 Academics 1st Engineer 5 years

3 Academics Chief Engineer 14.5 years

4 Academics Chief Engineer 10 years

________________________________________________

The weighting process resulted in the construction

of perception as cognitive function, where the task-

load has a relative importance of w

T=0.26, expertise

with w

E=0.41, and familiarity with wF=0.33. Similarly,

in the construction of comprehension, task-load had a

relative importance of w

T=0.13, expertise with wE=0.55,

and familiarity with w

F=0.32. Lastly, the weight is also

applied to determine the importance of each CF in

construction FM. In this case, perception had w

P=0.14,

while compression has w

C=0.86, in constructing the SA

loss as the main goal HRA in this study.

[ ]

0,1

= ∈

∑

j ij i j

D DW D

(1)

596

The relative importance result is used to define the

distance between the child node and its parent nodes.

The probability of the child node in certain states

should be assigned smaller if the parent nodes are in

different states. Take the example of CF perception

node, if its PSF task-load node is in the easy state, PSF

expertise node in high state and PSF familiarity node

in good state, the PSF node perception should have

higher probability of being in the high state compared

to medium and low states. Røed suggested the

conditional probability can be measured using this

distant methods [12]. Considering the direction of

change its parent nodes, we applied the Li method

that modified the equation into the absolute value, as

mention in the equation above, to distribute the

probability of the child node [16].

2.1.3 Conditional probability for child nodes

[ ]

0,1

−

= ∈

∑

(2)

To assign the probability distribution to each state

on child nodes, the formula from Røed is used [12]. In

the equation, the numerator is used to determine the

probability of each state of the child node in the focus,

while the denominator is normalization factor that all

state of the child sum up to 1. The higher R index will

make lower probability where the node in focus is

state distant from the parent’s states. While Røed uses

the expert input to decide the R index, this study

demonstrates how to reduce the uncertainty by using

the simulator result. Approach using simulator data

to decide R index has been demonstrated by Li in the

nuclear power plant analysis [16]. In this study, we

apply different approaches to decide the R index.

Table 2. Perception and comprehension sensitivity from

simulator experiment

________________________________________________

Variable Perception, Comprehension

________________________________________________

Task Load Easy Complex

________________________________________________

Familiarity Good Bad Good Bad

________________________________________________

P1 1.40 0.54 0.42 0.42 1.81 1.38 0.95 0.00

P2 1.38 0.42 1.40 0.95 1.94 0.42 0.00 0.00

P3 2.35 1.38 1.38 0.42 1.40 0.95 -0.42 0.00

P4 2.77 0.95 0.95 0.42 2.35 0.00 0.86 0.00

P5 2.35 0.42 1.38 1.40 1.40 0.54 0.97 0.42

P6 2.77 1.94 1.93 0.00 2.77 0.42 2.77 0.54

P7 2.35 0.95 1.38 0.95 0.97 1.81 1.93 -0.54

P8 2.77 0.42 1.81 0.95 0.97 0.00 1.40 0.95

P9 1.38 1.38 1.94 0.95 1.94 1.81 2.35 0.97

P10 1.40 0.54 0.95 0.42 1.81 0.00 -0.54 -0.95

P11 1.38 0.00 1.81 0.00 0.97 0.00 1.38 0.00

P12 1.40 0.42 1.38 0.95 1.81 0.42 0.54 0.00

P13 1.81 -0.54 0.97 -0.42 0.97 -0.42 -0.43 0.00

P14 2.77 0.00 0.95 -0.42 0.54 0.54 0.97 0.00

P15 1.94 0.00 1.94 -0.42 0.95 -0.42 0.86 0.00

P16 2.35 0.42 1.38 0.54 2.35 0.42 0.97 -0.95

________________________________________________

The simulator result is explained in Table 2. The

simulator experiment is designated with two task-

load states: easy where the simulation is under ocean

going scenario, and complex for entering-port

scenario where stand-by engine procedure must be

conducted. Familiarity has a bad state where the first

measurement is taken, and good when the repeated

measurement is taken. The sensitivity of perception

and comprehension level of situation awareness

measured by freeze-probe methods under signal

detection theory (SDT) [17,18]. The higher number on

it means the ability of each participant to discriminate

between the false alarm and correct rejection of the

questioned parameter is better. The participants P1

until P8 are categorized with expertise state high, and

P9 until P16 categorized with expertise state low.

Within this data, two thresholds are applied to

categorize the measurement result into three states:

low, mediate, and high. Important to note that the

average mediate state should have the larger number

of distributions.

Table 3. Simulator result distribution for perception

________________________________________________

Parent or CF Pr (CF | Parent State)

________________________________________________

Task-load Easy Complex

________________________________________________

Experience High Low High Low

________________________________________________

Familiarity Good Bad Good Bad Good Bad Good Bad

________________________________________________

Percep- High 0.75 0.00 0.25 0.00 0.25 0.13 0.13 0.13

tion Mediate 0.25 0.80 0.75 1.00 0.75 0.50 0.75 0.38

Low 0.00 0.13 0.00 0.00 0.00 0.38 0.13 0.50

________________________________________________

Table 4. Simulator result distribution for comprehension

________________________________________________

Parent or CF Pr (CF | Parent State)

________________________________________________

Task-load Easy Complex

________________________________________________

Experience High Low High Low

________________________________________________

Familiarity Good Bad Good Bad Good Bad Good Bad

________________________________________________

Compre- High 0.13 0.13 0.00 0.00 0.13 0.00 0.13 0.00

hension Mediate 0.88 0.75 0.50 0.50 0.63 0.38 0.38 0.13

Low 0.00 0.13 0.50 0.50 0.25 0.63 0.50 0.88

________________________________________________

The assigned conditional probability retrieved

from the simulator result was then compared with the

conditional probability from the Bayesian network

with R value varied from 0 to 5. We simply assign

RMSE func

tion as stated below to decide the

comparison between probability from simulator data

e and the probability from Bayesian model f with

scenario i and the state j. The m and n explains the

numbers of scenario and number states respectively.

The R value was then decided based on the

comparison which has the lower RMSE. In this case,

R=3.15 and R=1.22 are assigned for calculating with

Equation 2 to calculate the CPD for perception and

comprehension nodes. The result of each CPD is

shown in Table 5 and Table 6 for the perception node

and comprehension node, respectively.

( )

= −

∑∑

ij ij

RMSE e f

(3)

Table 5. Perception CPD

________________________________________________

Parent or CF Pr (CF | Parent State)

________________________________________________

Task-load Easy Complex

________________________________________________

Experience High Low High Low

________________________________________________

Familiarity Good Bad Good Bad Good Bad Good Bad

________________________________________________

Percep- High 0.80 0.34 0.23 0.10 0.44 0.13 0.11 0.03

tion Mediate 0.17 0.55 0.64 0.46 0.46 0.64 0.55 0.17

Low 0.03 0.11 0.13 0.44 0.10 0.23 0.34 0.80

________________________________________________

597

Table 6. Comprehension CPD

________________________________________________

Parent or CF Pr (CF | Parent State)

________________________________________________

Task-load Easy Complex

________________________________________________

Experience High Low High Low

________________________________________________

Familiarity Good Bad Good Bad Good Bad Good Bad

________________________________________________

Compre- High 0.54 0.36 0.25 0.19 0.46 0.29 0.23 0.16

hension Mediate 0.30 0.42 0.46 0.35 0.35 0.46 0.42 0.30

Low 0.16 0.23 0.29 0.46 0.19 0.25 0.36 0.54

________________________________________________

2.1.4 Calculating target node probability

The SA loss as the failure mode in the proposed

Bayesian model in this study is binary state; means

either the condition meets the true or false state. While

the node in the cognitive function and performance

shaping factor level is constructed by three states, the

following step is necessary to align it with the failure

mode level. The condition probability distribution in

the cognitive function is assigned as equation below.

1= =

∑∑

j basis i ik ik

i ka

P P w PQ

(4)

The conditional probability of failure mode node P

is calculated based on the probability of each parent

cognitive function node P

ik with states k=a,b,c. The

weighted value w

i is retrieved from subject-matter

experts like the previous step in weighting the PSF

nodes. Q

ik is the corresponding adjustment for the

basis. We follow the practice by Li to use Pbasis=0.01 as

the basic probability of error in SA, and 100-fold as

the number of compromise adjustment. This

configuration will made the adjustment factor for

parent nodes state to have Q

ik=0.01, Qik=1, Qik=100 for

parent node high, mediate, and low respectively.

Given the example, when the Task-load is

Complex, Experience is high, and familiarity is bad,

the probability distribution of perception is 0.13,

0.64, and 0.23 from Table 5, and the probability

distribution comprehension is 0.29, 0.46, and 0.25.

Where the weighting can be retrieved from the

previous explanation of subject matter experts. The

probability of SA loss given this condition can be

calculated as follow:

( ) ( )

( )

0.14 0.13 0.01 0.64 1 0.23 100 0.86 0.29 0.01 0.46 1 0.25 100

0.253

= × ×+×+×+× ×+×+× =

fail basis

Figure 4. SA loss probability comparison from Bayesian

network and expert expectation

The result from the Bayesian network modelling is

shown in Figure 4. The comparison includes the

expert expectation that was measured using the free

scale on the paper. The expert given the condition of

simulator result and asked how likely the participant

will fail in attaining the SA in each scenario. The

validation is not with an aim to validate until the level

of unit, but the tendency of the pattern comparison. It

can be accepted the Bayesian network result is follow

the expert expectation pattern at the most, except in

interception of scenario familiarity is good and

expertise is low, the Bayesian network result have

tendency to have higher probability in these scenarios.

2.2 Model updating

( ) ( )

( )

| =

P DH

PHD PH

(5)

In this model updating method, the calculated

probability for each combination of performance-

shaping factor, cognitive function, and failure mode

will be re-calculated given the condition if the new

simulator data exists. The method is based on the

Bayesian inference as explained in the equation

below. The aim is to update the posterior distribution

( )

|P HD

with the prior known information

distribution

( )

with collected data from

likelihood model

( )

|P DH

that normalized with

probability of distribution

( )

The Bayesian

inference able to use every time the new data exists,

means the posterior data from one modelling process

became prior data for the next modelling study or

stage.

2.2.1 Prior distribution construction

Regarding SA loss as the one factor for human

error event, it can be explained using the binomial

distribution explained by equation below. The

distribution expresses the uncertainty about the

number of failures x occurred in the given condition

of demands n with the parameter probability of

failure p. This parameter p is uncertain that sometimes

derived from expert judgment or data. Groth et al

suggest the p can be retrieved by using the Bayesian

inference [19].

( ) ( )

−



= −





f xp p p

(6)

We used the same approach as the previous

method in deciding the p by using the beta

distribution. Probability density function as shown in

below equation express the beta function

(

)

αβ

with function to normalize the distribution. From

with explanation from Groth et al, the distribution can

be conjugated with binomial distribution, where

posterior distribution parameters

post and

post are

assigned using equation below [19]. The next step in

using Bayesian inference is with assigning the value

and

598

(

)

( )

(

)

αβ

−

(7)

= +

post prior

aa x

ββ

= +−

post prior

(8)

To specify the prior distribution p

0, we applied the

constrained non-information (CNI) distribution [19]

As shown in the following equation, the a is estimated

as the number of failure contained in the prior

distribution, and the denominator (

) is considered

as the number of demands. The beta distribution is

constructed by

=0.5 and

derived from the equation

constraint. The extract prior distribution p

0 from the

new existing simulator data is shown in Table 7.

( )

(

)

( )

αβ

= =

E Beta E p

(9)

Table 7. Prior distribution (p0)

________________________________________________

Expertise Task-load Familiarity E(p0)   Prior Distribution

________________________________________________

High Easy Good 0.145 0.5 2.9 p0, Beta (0.5, 2.9)

Bad 0.215 0.5 1.8 p

0, Beta (0.5, 1.8)

Complex Good 0.179 0.5 2.3 p

0, Beta (0.5, 2.3)

Bad 0.253 0.5 1.5 p

0, Beta (0.5, 1.5)

Low Easy Good 0.269 0.5 1.4 p

0, Beta (0.5, 1.4)

Bad 0.465 0.5 0.6 p

0, Beta (0.5, 0.6)

Complex Good 0.357 0.5 0.9 p

0, Beta (0.5, 0.9)

Bad 0.584 0.5 0.4 p

0, Beta (0.5, 0.4)

________________________________________________

2.2.2 Posterior distribution updating

Table 8. Simulator result distribution

________________________________________________

Variable Perception, Comprehension

________________________________________________

Task Load Easy Complex

________________________________________________

Familiarity Good Bad Good Bad

________________________________________________

P17 1.81 1.38 1.81 0.54 0.95 0.00 2.35 0.95

P18 1.81 0.42 1.40 0.00 0.43 0.00 0.95 0.00

P19 2.35 0.95 2.35 0.95 1.40 0.54 1.40 0.42

P20 1.94 0.54 1.81 0.54 1.40 0.00 2.35 0.54

P21 1.81 0.43 1.94 1.38 1.40 0.97 1.81 0.97

P22 2.35 1.38 2.35 2.35 1.40 0.97 1.38 1.38

P23 1.81 2.35 1.38 0.42 2.35 0.42 0.00 1.40

P24 2.35 1.38 2.35 0.95 2.35 1.38 1.81 0.42

________________________________________________

The process updating the prior distribution with

new simulator data can be obtained by utilizing

Equation 8. The aims is calculating the posterior

distribution p

1 based on the prior distribution p0 and

the new distribution parameters

post and

post. For

thus, number of failures x occurred in the given

condition of demands n need to be defined. Table 8

illustrates the additional simulator result from eight

participants with expertise level assumed to be high.

In this step, a different approach is used to highlight

the participant who has perception or comprehension

equal or below 0 are categorized as loss in SA. Based

on the categorization, as also shown on Table 9, there

are eight trials for each scenario combination, then

number of opportunities can be assigned n=8, and the

number of failure x is assigned in each scenario

combination. The

post and

post is retrieved using

Equation 8, and the updated SA loss for selected

scenario.

Table 9. Posterior distribution for (p_1)

________________________________________________

Exper- Task- Famili- Data  ` Posterior E(p1)

tise load arity (x/n) Distribution

________________________________________________

High Easy Good 0/8 0.5 10.9 p1,

(0.5, 10.9) 0.044

Bad 1/8 1.5 8.8 p

(1.5, 8.8) 0.145

Complex Good 3/8 3.5 7.3 p

(3.5, 7.3) 0.324

Bad 2/8 2.5 6.4 p

(2.5, 6.4) 0.282

________________________________________________

3 DISCUSSION

HRA methods to measure and predict human error in

maritime operation are already developed with

various methods. The Bayesian network and Bayesian

inference is used in this study to demonstrate method

which use the simulator data for the probability

distribution calculation. The Bayesian network has the

advantage of treating the dependencies of PSF, which

often treat independent of each other in recent HRA

methods. Further, Bayesian inference concept in the

second stage demonstrated another possibility to

update the probability distribution of human error if

the new data from the simulator exists. This has an

advantage since there is no necessity to reconstruct

the model.

In the first stage of model construction, the three-

node level is introduced. FM were introduced as the

possible process or activity that support event led to

the human error. Situation awareness is introduced as

single FM in this study. CF is introducing as human

mechanism in construct the activity, its assigned as

dependent variable in simulator experiment.

Perception and comprehension were introduced in

constructing situation awareness. Last, PSF is

introduced as the lowest node in the model. In

relation to the simulator experiment, the PSF

represents the independent variable. Three PSF are

assigned in the model: task-load, expertise, and

familiarity. The proposed methods to replace the

expert judgment in deciding R-value are

demonstrated to reduce the subjective expert

judgment uncertainty. However, the input from

expert judgment is still mandatory to put the

weighting factors between the nodes in the same level

node. The second stage demonstrates updating the

prior distribution with the Bayesian inference

methods. Here the additional simulator data is used

to recalculate it into posterior distribution.

The human error probability as the output from

the model was compared to the expert expectation for

each scenario combination. It is observed that

Bayesian network results follow the same pattern as

the expert expectation input. However, several

comparisons such the scenario with familiarity is

good and expertise is low, is have higher evaluation

from the Bayesian Network. Thus, the Bayesian

network in this model still lacks sensitivity, especially

in the scenario which has close result of human error

probability.

The general construction of HRA with Bayesian

network offers the flexibility to cover more PSF into

the model. However, considering the PSF that can be

observed in the simulator will limit to detect all PSF

exist in other studies. This is the coming limitation of

the Bayesian network in HRA context. Thus, the

incomplete representation cannot explain the complex

599

relationship between variables. The second limitation

from employing simulator data is, during the session,

the participant is well now their performance is being

observed. This implies the participants put more

effort during the simulation. This must be noted since

human error calculation is aimed to measure the error

during the normal condition.

Work onboard a ship is divided into big portions

of navigation and engine operation work. Defining

general PSF that includes the two areas is the future

challenge that must be considered. The remaining

challenge in HRA is the various definition of the

denominator in probability, the number of

opportunities for human error which remain wide

interpretation for each method. An approach to

combine the result of HRA needs an adjustment

method to tackle this challenge. Similar to the

reliability in hardware, such study may become

reference under the IMO guideline to include in the

formal safety assessment.

REFERENCE

[1] Wróbel K. Searching for the origins of the myth: 80%

human error impact on maritime safety. Reliability

Engineering & System Safety 2021;216:107942.

https://doi.org/10.1016/j.ress.2021.107942.

[2] Grech MR, Horberry T, Smith A. Human Error in

Maritime Operations: Analyses of Accident Reports

Using the Leximancer Tool. Proceedings of the Human

Factors and Ergonomics Society Annual Meeting

2002;46:1718–21.

https://doi.org/10.1177/154193120204601906.

[3] Cordon JR, Mestre JM, Walliser J. Human factors in

seafaring: The role of situation awareness. Safety Science

2017;93:256–65. https://doi.org/10.1016/j.ssci.2016.12.018.

[4] International Maritime Organization. International

convention on standards of training, certification and

watchkeeping for seafarers. London: International

Maritime Organization; 2011.

[5] Patriarca R. Human reliability analysis: Exploring the

intellectual structure of a research field. Reliability

Engineering and System Safety 2020;203:27.

https://doi.org/10.1016/j.ress.2020.107102.

[6] International Maritime Organization. Revised Guidelines

For Formal Safety Assessment (FSA) For Use in The IMO

Rule-making Process. 2018.

[7] Yang ZL, Bonsall S, Wall A, Wang J, Usman M. A

modified CREAM to human reliability quantification in

marine engineering. Ocean Engineering 2013;58:293–303.

https://doi.org/10.1016/j.oceaneng.2012.11.003.

[8] Liu J, Aydin M, Akyuz E, Arslan O, Uflaz E, Kurt RE, et

al. Prediction of human–machine interface (HMI)

operational errors for maritime autonomous surface

ships (MASS). J Mar Sci Technol 2021.

https://doi.org/10.1007/s00773-021-00834-w.

[9] Ahn SI, Kurt RE, Turan O. The hybrid method combined

STPA and SLIM to assess the reliability of the human

interaction system to the emergency shutdown system of

LNG ship-to-ship bunkering. Ocean Engineering

2022;265:112643.

https://doi.org/10.1016/j.oceaneng.2022.112643.

[10] Mkrtchyan L, Podofillini L, Dang VN. Bayesian belief

networks for human reliabil

ity analysis: A review of

applications and gaps. Reliability Engineering & System

Safety 2015;139:1–16.

https://doi.org/10.1016/j.ress.2015.02.006.

[11] Park J, Jung W, Kim J. Inter-relationships between

performance shaping factors for human reliability

analysis of nuclear power plants. Nuclear Engineering

and Technology 2020;52:87–100.

https://doi.org/10.1016/j.net.2019.07.004.

[12] Røed W, Mosleh A, Vinnem JE, Aven T. On the use of

the hybrid causal logic method in offshore risk analysis.

Reliability Engineering & System Safety 2009;94:445–55.

https://doi.org/10.1016/j.ress.2008.04.003.

[13] Vinnem JE, Bye R, Gran BA, Kongsvik T, Nyheim OM,

Okstad EH, et al. Risk modelling of maintenance work

on major process equipment on offshore petroleum

installations. Journal of Loss Prevention in the Process

Industries 2012;25:274–92.

https://doi.org/10.1016/j.jlp.2011.11.001.

[14] Aydin M, Akyuz E, Turan O, Arslan O. Validation of

risk analysis for ship collision in narrow waters by using

fuzzy Bayesian networks approach. Ocean Engineering

2021;231:108973.

https://doi.org/10.1016/j.oceaneng.2021.108973.

[15] Musharraf M, Bradbury-Squires D, Khan F, Veitch B,

MacKinnon S, Imtiaz S. A virtual experimental

technique for data collection for a Bayesian network

approach to human reliability analysis. Reliability

Engineering & System Safety 2014;132:1–8.

https://doi.org/10.1016/j.ress.2014.06.016.

[16] Li P, Zhang L, Dai L, Li X-F. Study on operator’s SA

reliability in digital NPPs. Part 3: A quantitative

assessment method. Annals of Nuclear Energy

2017;109:82–91.

https://doi.org/10.1016/j.anucene.2017.05.019.

[17] Nizar AM, Miwa T, Uchida M. Measurement of

situation awareness in engine control room: approach

for non-technical skill assessment in engine resource

management. WMU J Marit Affairs 2022;21:401–19.

https://doi.org/10.1007/s13437-022-00270-6.

[18] Hogg DN, Folles⊘ K, Strand-Volden F, Torralba B.

Development of a situation awareness measure to

evaluate advanced alarm systems in nuclear power plant

control rooms. Ergonomics 1995;38:2394–413.

https://doi.org/10.1080/00140139508925275.

[19] Groth KM, Smith CL, Swiler LP. A Bayesian method for

using simulator data to enhance human error

probabilities assigned by existing HRA methods.

Reliability Engineering & System Safety 2014;128:32–40.

https://doi.org/10.1016/j.ress.2014.03.010.