873

1 INTRODUCTION

The shipping industry has an essential role in global

trade and commerce, as the majority of goods are

transported by ship. Given the complexity of shipping

operations, ensuring the safety and efficiency of ships

and their crew is of paramount importance. Shipping

operation is understood as a complex socio-technical

system. It requires a complex interaction between

social and technical components to achieve the

intended goals. A complex socio-technical system has

unique properties called emergence phenomena. This

idea explained the situation where the system has

ability beyond its individual component ability when

it is working as a whole. Based on this perspective,

safety in ship operation is acknowledged as an

emergence phenomena arising from the complex

socio-technical system rather than a property of the

system [1].

Resilience engineering [2] is a core representative

of this new safety idea. Along with this initiative, the

Functional Resonance Analysis Method (FRAM) is a

well-established framework that has been introduced

to provide resilience engineering ideas in its

application. This method analyses the interaction

between different elements of a system and

understands how their dependency contributes to

system performance. The FRAM has been successfully

applied in a variety of domains, including safety-

critical industries such as aviation[3], offshore oil and

gas [4], and the maritime industry [5], [6].

The process of ship maneuvering is dynamic with

continuous command and feedback between the

Modelling Ship Officer Performance Variability Using

Functional Resonance Analysis Method and Dynamic

Bayesian Network

I.G.M.S

. Adhita, M. Fuchi, T. Konishi & S. Fujimoto

Kobe University, Kobe, Japan

ABSTRACT: Ship maneuvering is a complex operation with inherent uncertainties. To express this complexity

in system performance during the navigation process, an analysis model has been developed using Functional

Resonance Analysis Method (FRAM) and Dynamic Bayesian Network (DBN). The functional level of dynamic

work onboard is assessed and modeled using FRAM qualitatively, in which a key function and the function’s

potential coupling for specific instantiation are identified. Further analysis is done by integrating the FRAM

analysis with DBN for quantification. The evolution of system performance over time is determined through

changes in the probability of function’s mode, namely strategic, tactical opportunistic, and scrambled. The

model presented in this study concerns the fluctuation of ship officer performance to overcome the obstacles

during the encounter event. As a result, the integration of FRAM-DBN shows promising usability to evaluate

human performance. The essence of human adaptive capacity is also highlighted through system resilience

potency, that is, the potency to learn, respond, monitor, and anticipate. We also discuss how this finding

contributes to enhance safety analysis, in specific, to provide explicit representation of the dynamic in human

performance in ship navigation based on Safety-II idea.

http://www.tra

nsnav.eu

the

International Journal

on Marine Navigation

and Safety of Sea Transportation

Volume 17

Number 4

December 2023

DOI: 10.12716/1001.17.04.

874

officer on watch (OOW) and the helmsman. It

escalates in different traffic and environmental

situation. Indeed, the system hardly relays on the

OOW’s capacity to decide on appropriate action. This

implies the necessity of proper performance

adjustment to maintain the system works normally

under dynamic working condition. In this case,

human adaptability and flexibility are essential to

achieve this kind of purpose. The definition of safety

in this manner is introduced by a term called Safety-II

[7]. Hence, under this consideration, it is important to

understand deeper this adaptable performance more,

in specific of how the fluctuation affects system

output.

The application of FRAM in this study is to

identify the functional basis of ship maneuvering

activity and provide a systematic expression of it

through functions and potential couplings. Analysis

in FRAM is qualitative in nature. It provides an

explicit understanding of the system’s functionality in

the form of FRAM model but lacks in function’s

performance representation. One solution that can be

applied is cooperating with the method of

quantitative analysis. Some available studies have

existed regarding this matter, such as the application

of Monte Carlo evolution [3] and modified Fuzzy

FRAM-CREAM with cellular automata simulation [8].

Therefore, in addition to the FRAM, this study

employs a Dynamic Bayesian Network (DBN) for

mathematical modeling. DBN is an extended version

of the Bayesian Network (BN). It is a type of

probabilistic graphical model that can be used to

model complex, dynamic systems over time. BN has

been widely applied in the maritime industry,

especially to predict the probability of ship accidents

[9]–[11]. One of the advantages of DBN is this method

can handle temporal dependencies and allow for the

modeling of time-varying influences on system

behavior. The basic idea of this quantitative

expression is to model the discrete probabilistic

dependencies between functions at each point in time,

and then to propagate these dependencies over time

to express a dynamic change in the system.

A case study from ship handling simulation has

been chosen to perform the analysis. The FRAM-DBN

integration in this study aimed to develop a

comprehensive model of ship officer variability

performance, providing insights into how changes in

officer performance over time influence system

output. The meaning of using this specific case is to

provide actual-time segregation for every decision

initiative and actual evidence of performance for DBN

analysis. Furthermore, the FRAM model can also be

built for a specific instantiation. This is a simple

analysis that design to present what to expect from

performance adjustment, and how can normal

performance be disrupted and then produce

undesired outcomes. In conclusion, the essence of

human performance for establishing an adapted

system can be addressed, and what strategy must be

built to enhance it to possess a higher level of

resilience.

2 METHOD

This section explains a set of methods applied in this

study. First is the qualitative analysis using FRAM

and build the FRAM model. Second is integrating the

FRAM analysis with quantitative analysis based on

DBN. In addition to that method integration, the

control model from Cognitive Reliability and Error

Analysis Method (CREAM) has been brought to

provide characteristic for FRAM function. This

characteristic includes strategic, tactical,

opportunistic, and scrambled.

2.1 FRAM as a retrospective analysis

The FRAM [12], [13] has been widely implemented to

assess system safety in various fields. System

resilience [2] terminology, as a core of FRAM,

promotes new ideas for safety research by

acknowledging safety as an emergent phenomenon

instead of a property of the system. This implies the

need on looking for what was done in the everyday

operation of the system (Work-As-Done) and how

safety is present in the system. The term function is

used to express the need for something to be done by

the system. The function is classified into three

categories including human, technological, and

organization. FRAM has four basic principles,

including the equivalent of success and failure,

approximate adjustment, the principle of emergence,

and functional resonance.

Function in FRAM is presented as a hexagon with

six aspects to characterize the type of information in

function’s dependency. Input represents the

information or material to trigger the process in

function. Precondition refers to the conditions that

exist prior to the function starting its operation. In this

sense, precondition is complementary information

that explains the pre-event or preparation that has

been done before a function is carried out. Time is the

temporal dimension of the function, including the

timing and sequencing of events. Control represents

the mechanisms used to control the function,

including procedures, rules, and policies. Resource is

something that is used or consumed by the function,

including energy, competence, people, tools,

materials, etc. The output represents the results of the

function’s work. Outputs can be seen as a signal that

starts a downstream function.

The relationship of function is described in two

forms. First is temporal relationships, namely

upstream and downstream functions. This expresses a

function’s dependency on specific time observation.

The downstream function is affecting the process that

happened in the downstream function, and this role

can change over time based on the potential coupling

and time of function activation. Second, the general

role of background and foreground functions.

Background function presents a function that only has

output or input as its aspect. It means the function is

only affecting the other functions in the system. On

the other hand, the foreground function is a function

that has more than 1 aspect. This function receives

information from background or foreground

functions, and also produces an output for other

foreground or background functions. In other words,

875

these roles refer to the relative importance of the

function in the model. The foreground function

denotes the matter being studied in the model, i.e., the

focus of the investigation.

The essence of ship maneuvering events is derived

by conducting an experiment in ship handling

simulators. Furthermore, one specific result of the

simulation has been chosen as factual information to

build the FRAM model. The selected case is a case

where the participants experience a collision during

the simulation. In this case, FRAM is used as a

retrospective analysis. The identification is focused on

how the system should have functioned for achieving

its goal, which is to avoid the target ship safely. As a

result, the change of situation from normal

performance to disrupted performance can be

recognized.

2.2 Officer performance quantitative representation using

DBN

DBN is an extension of the Bayesian Network (BN)

with the ability to handle temporal dependency

among nodes that change over time. This advantage

makes DBN suitable to be applied for establishing an

explicit representation of the dynamic in performance

of ship officers during encounter events. DBN

presents a pair of time slices of BN (Xt-1 → Xt), where

Xt-1 is the initial BN that defines the initial probability

of P(X), and Xt is the BN in the next time slice that

contains the conditional probability of two-time slice.

This state transition probability can be expressed as:

( )

(

)

(| ) |

tt t t

P X X P X Pa X

−

∏

(1)

where P(X) is the set of variables;

is the i-th node

of time slice t;

( )

Pa X

is the parent node of

; nt is

the number of nodes in the t-th time slice. In order to

solve this mathematical equation, the SMILE modeler

provided by BAYESFUSION has been applied in this

work. In addition, the DBN modeling has also been

done using the GeNIe software.

Figure 1. Relationship between control mode and common

performance condition

In this study, DBN is used to provide a

quantitative expression of the qualitative analysis that

has been provided by FRAM. To integrate the FRAM

with DBN, it is necessary to define a characteristic for

each function to generate a conditional probability

table (CPT). In this study, we decided to use the

control mode expressed in CREAM [14] to generally

characterize the FRAM function. It consists of

strategic, tactical, opportunistic, and scrambled.

Strategic control mode involves considering the global

context, using a wider time horizon and higher-level

goals, leading to more efficient and robust

performance, and planning based on the functional

dependencies between task steps. Tactical control

mode involves performance based on limited

planning, ad hoc needs, and frequently used

procedures that may seem rule-based due to context

or performance conditions. Opportunistic control

mode entails determining the next action based on

salient features of the current context, frequently

resulting in functional fixation, driven by perceptually

dominant features of the interface or frequently used,

familiar heuristics. Scrambled control mode involves

unpredictable decision-making without much

thought, often occurring during high task demands or

in unfamiliar, rapidly changing situations that lead to

a loss of situational awareness, potentially

culminating in momentary panic. Hence, the temporal

relationship of the function’s mode in which the

explicit representation of changes in performance over

time is presented through Bayesian thinking.

(a)

(b)

Figure 2. Discrete probability distribution of control mode

based on (a) CPC’s output variability [15] and (b)

probabilistic approach [16].

The determination of control mode in CREAM can

be derived by evaluating a defined criterion of a task.

However, in this study, we approach the

determination of the control mode in the form of a

discrete probability distribution. Figure 2 shows one

example of the discrete probability of the control

model in a given context of a common performance

condition. Based on this basis, it is possible to use this

idea to determine the CPT for the initial condition of

DBN. Therefore, the estimation of CPT for FRAM

function in maneuvering event is generated by

coopering discrete probability of CREAM control

876

mode with an adjustment based on the common

knowledge of officer performance in ship navigation.

3 CASE STUDY

A result from a ship handling simulation experiment

has been chosen to provide a factual story of ship

maneuvering events. The chosen simulation contains

an accident event in which the action taken by the

participant in each time slice is recognized. A deeper

understanding of participant’s decisions has also been

elaborated through structured interviews. The

participant is a licensed Officer with one experience

on board a ship as a Cadet. This is partial information

provided for the purpose of providing contextual

evidence for FRAM-DBN analysis. This simulation is

designed with a high level of maneuvering ability

such that the full potency of the participant to cope

with the situation can be observed.

Figure 3. Ship trajectory of the simulation result.

The event is last about six minutes. In this

situation, the target ship is moving from the northeast

and heading southwest. On the other hand, the

participant’s ship is initially heading southwest and

the final destination is in the southwest. The difficulty

is rising because the target’s speed is faster than the

participant’s speed. Therefore, he needs to maneuver

his ship to the destination and avoid the target ship

safely. The clarification of the participant’s decision to

overcome the situation is presented in Table 2. Five

questions were asked of the participant for further

understanding of his decision, including:

− When actually do you start to think to make this

decision?

− What information do you need before making this

decision?

− What makes you decide to take this action? Why

did you do it at that time?

− What was your strategy to avoid this target ship at

this moment? Do you have thought about that?

Please explain.

− What factors do you consider the most to decide

this action?

Table 1. Time step and participant’s explanation for each

decision.

________________________________________________

Step Recorded Description

time

________________________________________________

0 ±00:00 Start to perform an action. Monitoring the

target ship’s situation and building a

strategy to reach the destination.

1 ±00:40 An initial decision has been made by

ordering “Port 20.” After seeing the radar

and understanding the target situation, he

made a sharp turn to the portside to avoid

the target. The participant says, “I did not

think much because the target ship was an

overtaking ship and my ship was a

maintenance ship. I admit that I have had a

feeling of colliding at this moment.”

2 Up to Continue to monitor the situation.

±03:00

3 ±03:17 Realize that the first decision was bad. He

tried to overcome the situation by ordering

“starboard 10” but did not work well. The

participant says, ”At about 3:00, I was very

embarrassed as the target ship approached.

I could not think of anything at that

moment. In fact, before making this

decision, I should have asked the target

about her intention.”

4 ±05:41 The collision accident happened

________________________________________________

4 RESULTS

4.1 FRAM model and analysis

The implementation of FRAM in ship navigation to

assess the potency of system resilience has been done

by Adhita et al. [5]. For simplification, the dynamic

FRAM model for ship maneuvering during the

simulation experiment has been introduced as shown

in Figure 4. The model consists of five background

functions, which present the focus function for being

studied, including <To monitor (by OOW>, <To do

direct lookout>, <To watch electronic devices>, <To

decide action (make judgment)>, and <To control the

rudder/engine>. The temporal changes in the

function’s dependency and function’s updated role

over time have also been presented in the model.

Figure 4. Simplified dynamic FRAM model for ship

maneuvering in a simulation experiment.

This simple FRAM model of ship maneuvering

activity (to avoid the target ship) in Figure 4 shows

the intended functional process of system

performance. <To monitor (By OOW)>, <To do direct

877

lookout>, and <To watch electronic devices> plays an

important role for monitoring and learning processes.

<To monitor (By OOW)> produces initial information

such as strategy and expectation about the current

vicinity situation to activate <To do direct lookout>

and <To watch electronic devices>. Once the

information is collected, the system starts to decide

what kind of response should be done to cope with

the situation through <To decide action (make

judgment)>. Soon after that, the anticipation strategy

is start to produce through the connection of <To

decide action (make judgment)> and <To follow-up

monitor (by OOW>. Finally, the process of function

activation is repeated over time until the target ship

can be avoided.

This model shows the potency of functional

resonance can be triggered in any connection that

exists between functions. Furthermore, the continuous

process of function activation can increase the

emergence of functional resonance. The longest the

repetition, the higher the tendency for emerging

resonance. The case of collision accident explained in

Section 3 shows an example of how this resonance

phenomenon affects function performance. An early

signal of high variability performance from

monitoring and lookout functions was felt at around

00:40. This is probably the primary cause of the

amplifying effect that emerges in function <To decide

action (make a judgment)> then produces an

unwanted outcome, in this case, the order of

“starboard 20.” The last adjustment has also

performed “too late” in terms of timing, which

resulting the effort to maintain the system to work

normally cannot be achieved.

4.2 DBN for modeling dynamic performance in ship

maneuvering

To perform the DBN calculation, first, we convert the

FRAM model to be Dynamic Bayesian Network

model as presented in Figure 5. The continuous

expression is marked by the edge with a number [1]

on it. This edge expresses that the information in <To

monitor (by OOW> at t=1 is updated by itself at t=0

and <To decide action (make judgment)> at t=0. Then,

the process is looped exactly as how it expresses in the

FRAM model. The specific time step (t=0 to t=4) to

iterate the calculation is stated based on the case study

in Chapter 3.

Figure 5. DBN model using GeNIe Software.

The number of CPTs in the initial node set

depends on the number of nodes’ characteristics and

the number of edges pointing to the node. In total, 176

combinations of conditional probability of function

have been created. As explained in Chapter 2, the

generated CPTs are determined using the adjusted

description of control mode with common knowledge

of ship navigation. For example, <To do direct

lookout> and <To watch electronic devices> are

treated equally, in which the function is

complementing each other. In the case of one function

performing “strategic”, and the other performing

“scrambled”, it will affect <To decide action (make

judgement)> output more likely to be “tactical” or

“opportunistic” as presented in Table 2. The logical

way of thinking is the same as using “if-then rules” in

an intuitive way.

Table 2. Example of discrete CPT for <To decide action>.

________________________________________________

To do Strategic

direct

lookout

________________________________________________

To watch St Ta Op Sc

electronic

devices

________________________________________________

To Strategic 0.9558 0.45 0.0511 0.003

decide Tactical 0.0442 0.545 0.6333 0.4226

action Opportunistic 0 0.005 0.3138 0.4854

Scrambled 0 0 0.0018 0.089

________________________________________________

St – Strategic

Ta – Tactical

Op – Opportunistic

Sc – Scrambled

The adjusted value of CPT is also considering the

situation being assessed. It includes the familiarity

with the situation in the simulation, the difficulty

level of the encounter event given the participant’s

personal experience, etc. This considers important to

produce reasonable results. The proposed DBN is

intentionally exclusive for the case study.

Table 3. Example of discrete CPT for <To do direct lookout>.

________________________________________________

To monitor (by OOW)

________________________________________________

To do Strategic Tactical Opportunistic Scrambled

direct

lookout

________________________________________________

Strategic 0.75 0.05 0 0

Tactical 0.25 0.8 0.1 0

Opportunistic 0 0.15 0.85 0.2

Scrambled 0 0 0.05 0.8

________________________________________________

As a result, Figure 6 shows the expected

performance of the officer to maneuver the ship in the

encounter event. The probability of the performance

to be strategic is around 10% to 16%, tactical is around

32% to 48%, opportunistic is around 30% to 39%, and

scrambled is around 8% to 13%. The up and down in

each time slice shows the updated belief of

performance due to the function’s dependency, which

is quite stable and reasonable. This implies the

suitability of the given CPT and the model proposed

for the analysis has been achieved.

878

Figure 6. The evolution of expected normal performance of

ship officer in each time step.

FRAM analysis has provided qualitative analysis

for the case study. The fact that the emergence of a

scrambled mode of <To decide action (make

judgment)> at t=3 (±03:17) and speculation of the

potency of functional resonance as well as the

possibility of impact from the initial decision

possesses important information for the DBN

evaluation. Therefore, we proposed three

assumptions to present the situation in a more explicit

way.

Figure 7. The evolution of performance in each function

given the evidence of scrambled mode of <To decide action

(make judgment)> at t=3.

The first assumption: the fact that the scrambled

mode of <To decide action (make judgment)> has

emerged at t=3 (±03:17). Figure 7 present what is the

possible situation that happened before t=3 and how

the accident happened at t=4. In this case, the

evidence of <To decide action (make judgment)> is set

to be 100% scrambled. It can be seen, the probability

of scrambled and opportunistic modes before t=3 is

increasing in all functions. Specifically, <To monitor

(by OOW)> shows the worst situation among others.

In addition, the probability of <To control rudder> to

be in a scrambled mode is increasing up to more than

50% presents how bad the decision at t=3 was so that

the accident happened at t=4.

(a)

(b)

Figure 8. The evolution of performance in each function

given the evidence of scrambled mode of <To decide action

(make judgment)> at (a) t=1 and (B) t=3.

(a)

(b)

Figure 9. The evolution of performance in each function

given the evidence of scrambled and opportunistic modes of

<To decide action (make judgment)> at (a) t=1 and (B) t=3.

The second assumption: the first assumption

shows that the extreme changes in <To decide action

(make judgment)> at t=4 has a strong tendency to

indicate huge changes happened in the previous time

step. It is strengthened by the participant's argument

about his first action which is not carefully

considering the target situation. Therefore, Figure 8

shows the situation if at t=1 <To decide action (make

judgment)> is in the scrambled mode and at t=3 the

same mode appeared for the second time. This

situation shows that all functions are turned into an

extremely bad situation.

879

The third assumption: the best scenario that was

expected to occur. Given the second assumption, if the

scrambled mode occurred at t=3, there is about 20% to

30% chance for the opportunistic mode to happen in

the next sequence of function activation. Let set <To

decide action (make judgment)> to be opportunistic at

t=3. The result in Figure 9 shows the system

drastically change to the opportunistic mode as

expected. This expresses how the situation could be if

the participant tries to communicate with the target

ship as he mentioned. In addition, this could also

express the possible adjustment if the action to

overcome the bad decision at t=1 was taken before

3:00.

5 DISCUSSION

This study presents the solution for assessing safety

based on the Safety-II perspective using FRAM-DBN

analysis. A case study of a ship collision accident has

been chosen to provide system degraded performance

from normal to disrupted. This collaboration has been

found excellent, especially to assess the dynamic

changes in function performance over time. The

proposed method is able to provide a further

understanding of the function's temporal dependency.

The DBN analysis complements the FRAM analysis

and provides a more in-depth understanding of the

function’s performance in the encounter event.

Moreover, the proposed DBN can be used as a

decision support tool for officers in similar situations,

providing insights into the expected performance and

potential consequences of different actions.

The analysis shows the essence of performance

adjustment for establishing safety in ship

maneuvering. The accident was found to happen due

to the inability of the system to produce a proper

adjustment. This is strengthened by the fact that the

target ship hits the stern side of the participant’s ship.

It indicates that a slightly better adjustment in the

decision at t=3 to overcome the unwanted

performance at t=1 could prevent the collision

accident to happen. Furthermore, the assumptions in

the DBN analysis proposed helped to illustrate these

potential consequences of different decisions made by

the officer, highlighting the importance of decision-

making skills in ship navigation.

The concept of Safety-II strongly encourages

approaching safety from how it is present in everyday

operations. Although the case being studied is the

accident event, the mean is to provide the point of

view of normal and disrupted performance, such that

both situations can be explicitly presented. The

importance of local adjustment in the ship

maneuvering process has been highlighted. From

different instance in different time slice, the officer

performs different strategy to continuously follows

the working dynamic. In this current example, the

unwanted adjustment is presented. However, the

understanding of what is expected could be can also

be provided. It indicates the need for enhancing

human performance flexibility for a better level of

ship resilience. Obviously, there is a boundary for

system flexibility to cope with a certain level of

dynamic situations. Finding a balance of it can be

another problem to solve. Given today’s phenomena

of AI and autonomy, a higher level of ship resilience

can also be achieved by incorporating humans and

technology through human-autonomous interaction.

For a more comprehensive analysis, future studies

must consider a modification in the input data to

determine CPT. Expanding the network can also be

considered for more understanding of specific factors

that influence the change in performance. The FRAM

also facilitates expandability, especially for the

“loose” couplings in function.

6 CONCLUSION

The use of FRAM-DBN analysis in this study provides

a valuable tool for analysing the performance of ship

officers during the maneuvering process. Continuous

expression of changes in officer performance over

time can be greatly presented using the dynamic

FRAM model and discrete probability distribution of

the function’s performance mode in DBN. This

elaborated FRAM analysis shows what is to be

expected in the normal performance of ship

navigation and how the performance degradation

happened based on the case study being assessed. The

application of this proposed method is limited to the

case being analysed in this research. However, the

usability for a more complex implementation has been

recognized. In order to enhance the resilience of ship

navigation, a thorough understanding of human

flexibility and adaptability in response to unexpected

situations is essential.

ACKNOWLEDGEMENT

We would like to acknowledge the Japan Society for the

Promotion of Science (JSPS KAKENHI Grant Number

22K04598) for funding this research. In addition, the DBN

models discussed in this paper were constructed with the

GeNIe Modeler, a tool freely available for academic research

and teaching use from BayesFusion, LLC.

REFERENCES

[1] E. Hollnagel, R. L. Wears, and J. Braithwaite, “From

Safety-I to Safety-II : A White Paper From Safety-I to

Safety-II : A White Paper Professor Erik Hollnagel

University of Southern Denmark , Institute for Regional

University of Florida Health Science Center Jacksonville

, United States of America Prof,” no. October, 2015.

[2] E. Hollnagel, D. D. Woods, and N. Leveson, Resilience

Engineering: Concepts and Precepts. Ashgate, 2006.

[3] R. Patriarca, G. Di Gravio, and F. Costantino, “A Monte

Carlo evolution of the Functional Resonance Analysis

Method (FRAM) to assess performance variability in

complex systems,” Saf. Sci., vol. 91, no. October, pp. 49–

60, 2017.

[4] J. E. M. França, E. Hollnagel, and G. Praetorius,

“Analysing the interactions and complexities of the

operations in the production area of an FPSO platform

using the functional resonance analysis method

(FRAM),” Arab. J. Geosci., vol. 15, no. 7, 2022.

[5] I. G. M. S. Adhita, M. FUCHI, T. KONISHI, and S.

FUJIMOTO, “Ship Navigation from a Safety-II

880

Perspective: A Case Study of Training-ship Operation in

Coastal Area,” Reliab. Eng. Syst. Saf., p. 109140, Feb.

2023.

[6] I. G. M. S. Adhita and M. Furusho, “Ship-to-Ship

Collision Analyses Based on Functional Resonance

Analysis Method,” J. ETA Marit. Sci., vol. 9, no. 2, pp.

102–109, 2021.

[7] E. Hollnagel, Safety-I and Safety-II: The Past and Future

of Safety Management. CRC Press, 2014.

[8] T. Hirose and T. Sawaragi, “Extended FRAM model

based on cellular automaton to clarify complexity of

socio-technical systems and improve their safety,” Saf.

Sci., vol. 123, no. November 2019, p. 104556, 2020.

[9] M. Hänninen and P. Kujala, “The effects of causation

probability on the ship collision statistics in the Gulf of

Finland,” Mar. Navig. Saf. Sea Transp., vol. 4, no. 1, pp.

267–272, 2009.

[10] Q. Yu and K. Liu, “An expert elicitation analysis for

vessel allision risk near the offshore wind farm by using

fuzzy rulebased bayesian network,” TransNav, vol. 13,

no. 4, pp. 831–837, 2019.

[11] R. Billard, J. Smith, M. Masharraf, and B. Veitch, “Using

Bayesian networks to model competence of lifeboat

coxswains,” TransNav, vol. 14, no. 3, pp. 585–594, 2020.

[12] E. Hollnagel and Ö. Goteman, “The Functional

Resonance Accident Model,” Proc. Cogn. Syst. Eng.

Process plant, pp. 155–161, 2004.

[13] E. Hollnagel, FRAM: the Functional Resonance Analysis

Method. England: Ashgate, 2012.

[14] E. Hollnagel, Cognitive Reliability and Error Analysis

Method (CREAM), First Edit. Elsevier Ltd, 1998.

[15] T. Bedford, C. Bayley, and M. Revie, “Screening,

sensitivity, and uncertainty for the CREAM method of

Human Reliability Analysis,” Reliab. Eng. Syst. Saf., vol.

115, pp. 100–110, 2013.

[16] M. C. Kim, P. H. Seong, and E. Hollnagel, “A

probabilistic approach for determining the control mode

in CREAM,” Reliab. Eng. Syst. Saf., vol. 91, no. 2, pp.

191–199, 2006.