Accident Data-Driven Consequence Analysis in Maritime Industries

Abstract

Maritime accidents are significant obstacles to the development of shipping industries. Their consequences are another important issue because they often involve significant economic losses and human casualties. Accident consequences do not occur randomly, but are triggered by a series of influential factors. To determine the critical factors contributing to accident consequences, a data-driven research framework is proposed. Firstly, 198 maritime accident investigation reports from the Marine Accident Investigation Branch (MAIB) and Australian Transport Safety Bureau (ATSB) are collected to build a database. Secondly, relevant influential factors are identified based on a literature review. Thirdly, a TAN (Tree Augmented Network)-based BN (Bayesian network) model is developed. Fourthly, a model validation process, including a comparative analysis, Kappa test, and scenario analysis are performed. The five critical factors are determined as accident type, ship type, ship age, ship length and gross tonnage. Valuable implications are generated through this research framework and can be a valuable reference for the safety management of concerned parties. In addition, the TAN model can be a predictor for developing mitigation measures to minimize accident consequences.

1. Introduction

Although there have been significant efforts to control maritime accidents and their related consequences, the results seem to be unsatisfactory. Data [1] indicate that 27,477 accident records have been collected around the world in the last 10 years (from 2013 to 2022). Among the total 27,477 accidents, 807 led to catastrophic consequences (total loss), and the losses caused are undoubtedly enormous. Other organizations or agencies, such as EMSA (European Maritime Safety Agency) [2], DNV (Det Norske Veritas) [3], etc., have published similar reports. Thus, maritime accidents and their consequences are still major obstacles to maritime safety and the shipping industry. To improve maritime safety and minimize accident consequence costs, effective measures should be developed.

Compared to the intense level of research on maritime accidents, research on their consequences is relatively rare and uncommon. Siddiqui et al. [4] performed a study on oil spill consequences and found that the route length is not positively correlated to the consequence severity. Wu et al. [5] analyzed collision accident consequences with a BN model, using expert judgment to help build the BN structure. Except for the research on specific types of consequences, other kinds of consequence analysis have also been performed by scholars. Wu et al. [6] used maritime accident reports to develop an analysis model to ultimately study the mechanisms driving the consequences of maritime accidents, with expert decisions helping to develop the model. To help develop consequence mitigation measures, Li et al. [7] utilized a prediction approach to mine the risk influential factors (RIFs) contributing to accident consequences. Similarly, Zhang et al. [8] statistically analyzed the RIFs influencing accident consequences; 477 maritime accident reports were collected in their study, and the vessel speed and crew number were highlighted.

The MAIB is an organization independent from the British government [9]. The maritime accident investigation reports from MAIB include the statement: “The sole objective of the investigation of an accident shall be the prevention of future accidents through the ascertainment of its causes and circumstances. It shall not be the purpose of such an investigation to determine liability nor, except so far as is necessary to achieve its objective, to apportion blame”. Therefore, the data collected here are considered objective and thus suitable for this research. Similarly, ATSB is an organization independent from the Canadian government [10]. The accident reports from ATSB have the same characteristics as the MAIB reports. In this research, a database is intentionally developed by collecting the accident reports from MAIB and ATSB. In addition, the following three points are also the contributions of this research.

(1) Thanks to the objectivity of the data sources, this research can help improve the accuracy of accident severity prediction and consequence analysis when applied to the research process.

(2) Maritime accident consequence analysis suffers from data shortages [7]. The database developed here can bridge the research gap and stimulate data-driven accident consequence research in the future.

(3) This research explores the critical factors in maritime accident consequences through a newly developed data-driven BN model. The database’s objectivity can be transmitted to the results, which can guide the upgrade of maritime accident consequence management.

The rest of the paper is organized as follows. Section 2 reviews the related literature and finds the research gap. Section 3 gives the general methods of this research. Section 4 presents the detailed validation analysis, sensitivity analysis, comparative analysis and resulting implications. Finally, Section 5 presents the conclusions of this research.

2. Literature Review

2.1. RIFs of Maritime Accidents

Maritime accidents are triggered by various factors, and these factors have aroused considerable attention in academic study. Ma and Deng [11] developed a complex network approach to determine the RIFs of grounding accidents and found that the controlling strategies may be different according to the situation. Fan and Yang [12] developed a machine-learning model to examine maritime accidents in restricted waters. Weng et al. [13] chose a statistical method to assess maritime accidents worldwide and determine the accident mortality. Maya et al. [14] utilized a fuzzy cognitive map to assess the RIFs of maritime accidents. With the 24 model, Lyu et al. [15] proposed a method for smart safety management. Chen et al. [16] used an odds logistic model to analyze total-loss maritime accidents and found that foundering and collision are the main types. Wang et al. [17] found that a vessel’s safety conditions and technical characteristics are the most important RIFs influencing the PSC (Port State Control) inspection and detention. Gucma and Androjna [18] utilized a simplified evaluation framework to study maritime accidents. Li et al. [19] analyzed the related influential factors of maritime accidents through a BN model, and ship type and ship operation were considered the most correlated factors. Sevgili et al. [20] focused on oil spill accidents in oil tankers with a data-driven BN model, and accident type and vessel age. Chen et al. [21] conducted systematic research on global total-loss maritime accidents. Stranding and foundering were identified as the critical RIFs. Maya et al. [22] evaluated the accident mechanisms among oil tankers and identified the critical factors. Zhang et al. [23] analyzed the critical RIFs in collision accidents, proposing a hybrid model (BN, least square) and finding that small vessels are more prone to maritime accidents. Jiang et al. [24] found that the vessel’s speed, type, and age are the critical RIFs of maritime accidents for vessels trading in the Maritime Silk Road. Chen et al. [25] developed an analysis model to study human factors involved in maritime accidents. Yıldırım and Başar [26] outline the human factors influencing maritime accidents through the HFACS (Human Factor Analysis and Classification System) framework. Similarly, Paolo and Gianfranco [27] conducted a similar analysis but with the cluster analysis method. Uğurlu and Umut [28] utilized an evaluation approach to examine the human factors in grounding accidents. Kum et al. [29] used the marine accident reports from MAIB to analyze the significant RIFs and found that the crew’s carelessness is the root cause. Fu et al. [30] review the research of maritime safety in Arctic shipping and the RIFs of marine accidents were discussed and identified. Luo and Shin [31] carried out a systemic review of research of maritime accidents. Kulkarni et al. [32] review the research status of ship accident prevention in the Baltic Sea area.

2.2. Consequences of Maritime Accidents

The consequences of maritime accidents are not only affected by the accidents themselves, but also influenced by the mitigation measures, emergency handling skills, and certain environments, etc. Browne et al. [33] propose an approach to studying arctic ship accident severity. The significant RIFs are determined, and oil tankers and cruise ships are targeted as the most sensitive ship types. Siddiqui and Verma [4] propose an assessment method to perform oil spill accident consequence analysis, finding that the oil spill quantity is not positively correlated with route length. Pitblado et al. [34] analyzed the accident consequences of chemical tankers, highlighting the dangerous areas of tankers. Wu and Tian [5] built a data-driven model aimed at studying the collision accident consequences, and the related emergency handling is well addressed. Wu and Leung [6] assessed the maritime accident consequences with an evaluation approach, using 797 accident records to develop the model. Wang et al. [35] analyzed the RIFs of maritime accident consequences by gathering accident reports from the maritime safety administration of China. Zhang and Wang [8] statistically analyzed accident consequences and generated useful insights into relevant mitigation measures. Maritime accidents can result in very severe consequences, like the Titanic accident. Hence, it is critical to investigate their mechanisms and develop mitigation measures to narrow the research gap.

2.3. Research Gaps

Previous studies on marine accident consequences are limited and can be divided into two groups: analyses of specific consequences, and studies on the influence mechanisms of RIFs on accident consequences. However, among the models and methods used, data-driven BN (i.e., TAN learning) is relatively less common. One possible reason is the large quantity of data it requires, and the data are not easy to collect. The advantages of TAN-based BN are obvious compared to the above-mentioned methods: expert evaluation is not required, forward prediction is included, and backward diagnosis is possible. The flexibility of TAN learning makes it suitable for the analysis of maritime accident consequences. On the other hand, the Kappa statistic test is firstly applied to the TAN-trained BN model. The model is highly associated with the input database and thus is unique compared to other TAN learning generated models. Therefore, this research utilizes a TAN-based BN to uncover the interrelationships between RIFs and accident consequences.

3. Research Methods

3.1. The Framework of the Method

Figure 1 shows the framework of the method. First, the WOS (Web of Science) core collection was searched for research article collection, and the RIFs of maritime accidents in these articles were statistically analyzed to identify the high-frequency RIFs. Secondly, an expert evaluation was applied to check and adjust the selected RIFs, so that the identified RIFs could be determined. Thirdly, the accident reports from MAIB and ATSB were collected and refined to build a specific database. Fourthly, with the database, a data-driven BN was constructed, and the BN structure and parameter learning were performed. Fifthly, the model validation, including the sensitivity analysis and consistency analysis, were carried out. Finally, the implications are generated.

3.2. Data Sources

The maritime accident reports from MAIB and ATSB are open source data. These data can be downloaded directly from their official websites. However, not all the reports uploaded to the ATSB website as completed final editions; therefore, a screening process to identify valuable reports is necessary. Only final reports marked as ‘accident’, ‘incident’ and ‘serious incident’ from the period between January 2010 and December 2022 were collected. After this filtering process, 64 reports were collected. Similarly, 134 reports from the MAIB were collected. As a result, 198 reports were collected to form the database for this research. The distribution of the accident types in the database can be seen in Figure 2. ‘Accident to person’ holds the highest frequency of occurrence, and ‘grounding’ and ‘collision’ are located in the second and third positions. To provide a reasonable dataset for the model training, the collected dataset was intentionally divided into two sets at a ration of 8:2. The training dataset had 158 records while the test dataset had 40 records. The two datasets were used for model training and model validation, respectively.

3.3. Identifing and Determining the RIFs

Based on the keywords ‘maritime risk’, ‘maritime accident’ and ‘maritime consequence’, the Web of Science core collection was searched and 20 research articles (

Table 1. Collected articles.

NO	Citation	Focus	Method	Database
1.	[39]	Human factors, maritime accident	HFACS (Human Factors Analysis and Classification System), grounding theory, ARM (Association rule mining)	Accident reports from China MSA (Maritime Safety Administration of China), JTSB (Japan Transport Safety Board), and NTSB (United States National Transportation Safety Board)
2.	[24]	Risk, maritime accident	BN-based risk analysis model	Accident reports from LRF (Lloyd’s Register Fairplay), IMO database
3.	[21]	Risk, total-loss maritime accident	BN, TOPSIS (Technique for Preference by Similarity to Ideal Solution)	Collection of total-loss maritime accidents in the world from 1998 to 2018
4.	[36]	Accident prevention	Data-driven BN	MAIB and TSB (the Transportation Safety Board of Canada)
5.	[38]	Accidents causation	Data-driven BN, machine learning	Accident reports from China MSA
6.	[35]	Accident consequence, Chinese waters	ABN (augmented naïve BN) model	Accident reports from China MSA
7.	[5]	Yangtze River, accident consequence	BN	Historical accident data of Yangtze River
8.	[8]	Accident consequence, RIFs	Kendall’s tau coefficient, Cramer’s V analysis and Kruskal–Wallis test	Accident reports from EMSA (European Maritime Safety Agency), MAIB
9.	[6]	Yangtze River, collision accidents	BN	Historical accident data of Yangtze River
10.	[13]	Accident injury severity	Binary logistic regression model, zero-truncated binomial regression model	Accident reports from Lloyd’s List Intelligence
11.	[12]	Confined waters, maritime accident	Data-driven risk model	Accident reports from MAIB and TSB
12.	[40]	Istanbul strait, maritime accident	Data-driven BN	418 vessel accident records that occurred in the Istanbul Strait
13.	[41]	Accident consequence	Data-driven BN	Accident reports from China MSA, JTSB
14.	[42]	Maritime casualty, risk	Data-driven method	Accident reports from LRF, IMO database
15.	[43]	Risk management, Yangtze River	System-based model	Accident reports from China MSA
16.	[44]	Accident prediction	Machine learning	The maritime accident log data from January 2010 to December 2020 in Korean waters
17.	[45]	Yangtze River, MASS	Data-driven BN	Accident reports in the Yangtze River from China MSA
18.	[46]	Maritime accident	BN-based risk model	Accident reports from MAIB and TSB
19.	[47]	Maritime accident	BN-based risk model	Accident reports from the GISIS (Global Integrated Shipping Information System)
20.	[48]	Risk management, arctic waters	BN-based risk model	Accident reports from the GISIS, TSB

) were collected. With a statistical analysis of the gathered papers, the main RIFs involved were found and are shown in Figure 3. ‘Lookout’ holds the highest frequency of occurrence, followed by ‘ship type’, ‘ship age’, ‘gross tonnage’ and ‘visibility’. Referring to research [36,37,38,39], the average occurrence frequency (i.e., 7) was selected as the threshold to screen the RIFs for building the database and model construction. As a result, the selected RIFs were ‘ship type’, ‘lookout’, ‘ship age’, ‘ship flag’, ‘accident type’, ‘length’, ‘visibility’, ‘speed’, ‘gross tonnage’, ‘wind’, ‘traffic density’, ‘time of day’, and ‘equipment’. Finally, 13 RIFs were selected. Their states, listed in

Table 2. Defined RIFs.

NO.	RIFs	Description	State
1	Ship age [37]	(0 5], [6 10], [11 15], [16 20], >20, NA	1,2,3,4,5,6
2	Gross tonnage [21]	(0 3000], (3000 10,000], (10,000 20,000], >20,000	1,2,3,4
3	Ship length [21]	(0 100], (100 200], >200	1,2,3
4	Ship type [38]	Container ship, tanker, bulk carrier, passenger ship, general cargo ship, RORO, others	1,2,3,4,5,6,7
5	Equipment [6]	Good condition, available; bad condition, not available	1,2
6	Ship flag [4]	FOC *, non-FOC	1,2
7	Ship speed [39]	Safe speed, unsafe speed	1,2
8	Wind condition [21]	(0 5], >5	1,2
9	Visibility [40]	Good or bad	1,2
10	Time of the day [26]	Day (06:00/17:59), night (18:00/05:59)	1,2
11	Traffic density [38]	Low, high	1,2
12	Lookout [34]	Proper, improper	1,2
13	Accident type [49]	Collision, grounding, accident to person, fire/explosion, machinery failure, others	1,2,3,4,5,6
14	Consequence severity [50]	Minor, significant, severe, catastrophic	1,2,3,4

* FOC refers to flag of convenience. Please refer to [51].

, are given according to the collected 20 articles. ‘Consequence severity’ is the dependent node.

3.4. BN Modeling

Conventional BN usually builds its network structure and conditional probability tables through expert evaluation. The reliability of this process is arguable because of its subjectivity and potential bias. When the accident data are easier to obtain, the data-driven BN model starts to appear and shows better performance compared to the traditional models. Scholars have developed several methods for the application of data-driven BN, like TAN, k2 algorithm, ABN (Augmented Bayesian Network), and NBN (Naive Bayesian Network). Fan et al. [36] developed a TAN-BN approach to examine the marine accident from a human factors perspective. Based on ABN modeling, Wang and Yang [39] proposed an ABN model to investigate the relations between the accident consequence and RIFs. Ozaydın and Fıskın [52] incorporated BN and rule mining to study the accidents of fishing vessels. Several studies [53,54] have proven the TAN method shows better accuracy and adaptability compared to the other data-driven BN methods. Therefore, we utilize the TAN algorithm for the BN structure learning of this research.

The TAN method of BN modeling often involves four steps: collecting data, identifying the RIFs, building the BN structure, and validating the model. Our research also involved these four steps. The first two steps have been discussed in Section 3.2 and Section 3.3. The BN structure can be built with the help of the Netica app [55], utilizing the TAN algorithm for learning the BN structure and the gradient method for learning the conditional probability tables.

3.5. Model Evaluation

Considering the importance of the method’s justification, it is essential to evaluate and validate the proposed model. The validation process is executed through comparative analysis, a Kappa test, sensitivity analysis, and scenario analysis.

The comparative analysis can be achieved through the comparison between statistical results and the model predicted results. The Kappa statistic test is employed to calibrate the consistency between the two results. The test result is obtained with Equation (1).

$$t=\left(s_0-s_e\right)/\left(1-s_e\right)$$

(1)

where s₀ represents the accuracy of the two results, and s_e is the probability of consistency. The s₀ value is the ratio between the predicted value and the true value. The s_e value is the sum of the product of the predicted value and the true value divided by the square of the total number.

MI (Mutual information) analysis is the main form of sensitivity analysis because it signifies the independence between different nodes [45]. It can be defined in the context of BN modeling as follows:

$$M\left(c,f_i\right)=-{\displaystyle \sum _{i,j}P\left(c,f_{ij}\right)}{\log }_a{\displaystyle \frac{P\left(c,f_{ij}\right)}{P\left(c\right)P\left(f_{ij}\right)}}$$

(2)

where c represents “consequence severity”, f_i means the ith factor, f_ij is the jth state of the ith factor, and M(c,f_i) denotes MI. The M value implies the correlation strength between the two nodes. A larger value means a higher correlation.

TRI (True risk influence) serves as an alternative approach to sensitivity analysis, which has been proven in research [45]. The TRI value can be calculated after the HRI (high-risk influence) and LRI (low-risk inference) values are obtained. Reference [45] has outlined the calculation procedure.

4. Results and Discussion

4.1. Trained Model

Figure 4 shows the trained model of this research. It revealed that ‘accident to person’ (25.6%) is the highest accident type, followed by ‘others’, ‘grounding’, ‘collision’ and ‘fire/explosion’. These predicted results are actually in line with the statistical figures in Figure 2, which is the rational evidence of the trained model. The states of each node have their own occurrence probabilities, which can help analyze the mechanisms of maritime accident causation and provide useful insights for accident prevention.

4.2. Validation of BN Model

The validation process of BN modeling includes comparative analysis and the Kappa statistic test. The comparative analysis can be performed through the comparison between the predicted results and the statistical data.

Table 3. The comparison of ‘accident type’.

Accident Type	Statistical Results (%)	Predicted Results (%)	Difference (%)
Collision	18	18.1	0.6
Grounding	21	20.4	2.9
Accident to person	25	25.6	2.4
Fire/explosion	9	9.18	2
Machinery failure	6	5.83	2.8
Others	21	20.9	0.5

indicates the comparison of node ‘accident type’. Each of its states shows a slight difference: ‘grounding’, with 2.9% difference, has the biggest difference, followed by ‘machinery failure’ (2.8%), ‘accident to person’ (2.4%) and ‘fire/explosion’ (2%). Overall, all the differences are within 5%, which shows good agreement of the modeling results with the statistical data.

The test dataset built in Section 3.2 is then applied to the trained model, and a confusion matrix is generated, as demonstrated in

Table 4. The confusion matrix.

-	Minor	Significant	Severe	Catastrophic	Actual Total	Accuracy (%)
Minor	1	0	0	0	1	100
Significant	0	22	2	0	24	91.7
Severe	0	2	13	0	15	86.7
Catastrophic	0	0	0	0	0	100
Predicted total	1	24	15	0	40	90

. Then, the t value can be obtained with Equations (3)–(5).

$$s_e={\displaystyle \frac{\left(1\times 1+24\times 24+15\times 15+0\times 0\right)}{40\times 40}}=0.5103$$

(3)

k₀ = 0.9

(4)

t = 0.7958

(5)

Referring to research [52,53], when the t value is within the range of [0.61 0.8], it means a significant match when t ∈ [0.61 0.8]; and it shows an almost perfect match when t ∈ [0.81, 1]. The t value (0.7958) here indicates a good match of the trained result.

The sensitivity analysis, e.g., MI and TRI analysis, can help validate the proposed model and measure the correlations between RIFs and consequence severity.

Table 5. MI analysis results.

RIFs	MI	Percentage (%)	Variance of Belief
Accident type	0.16917	11.2	0.04089
Ship length	0.02834	1.87	0.00555
Visibility	0.02154	1.43	0.00154
Wind condition	0.02070	1.37	0.00073
Ship age	0.01898	1.26	0.00295
Ship flag	0.01702	1.13	0.00066
Gross tonnage	0.01681	1.11	0.00298
Traffic density	0.01325	0.876	0.00187
Ship type	0.01029	0.8	0.00160
Time of day	0.00665	0.44	0.00117
Ship speed	0.00543	0.359	0.00027
Equipment	0.00327	0.216	0.00051
lookout	0.00217	0.144	0.00008

shows the MI between RIFs and consequence severity, where the consequence severity is the dependent node. The MI of ‘accident type’ (0.16917) occupies the highest position, the second to the fourth positions are ‘ship length’, ‘visibility’ and ‘wind condition’.

Table 6. TRI values for gross tonnage and consequence severity.

(0 3000]	(3000 10,000]	(10,000 20,000]	>20,000	Minor	HRI	LRI	TRI
/	/	/	/	52.4	3.2	3.8	3.5
100%	0	0	0	48.6
0	100%	0	0	55.6
0	0	100%	0	53.5
0	0	0	100%	53.4

gives the TRI values between ‘gross tonnage’ and ‘minor’ consequence.

Table 7. All TRIs of this research.

	Minor	Significant	Severe	Catastrophic	Average
Accident type	5.6	22.6	28.25	2.66	14.78
Ship length	5.95	8.35	8.1	1.85	6.06
Visibility	4.47	8.1	0.05	4.66	4.32
Wind condition	2.57	1.75	4.55	3.73	3.15
Ship age	3.94	11.7	6.15	8.93	7.68
Ship flag	2.59	0.9	5.05	3.37	2.98
Gross tonnage	3.5	8.25	9.6	2.63	6.0
Traffic density	1.35	4.55	7.35	1.43	3.67
Ship type	3.04	5.8	8.4	1.51	9.38
Time of day	1.18	3.05	4.45	0.5	2.3
Ship speed	1.78	0	3	1.21	1.5
Equipment	1.01	1.85	2.95	0.1	1.48
Lookout	1.18	0.75	1	0.6	0.88

shows all TRI values of this research. According to the literature [45], a bigger TRI value indicates a closer relationship between the RIF and the accident consequence. Different levels of consequence severity are influenced by RIFs in various ways. Each severity level has its own most important RIF. For example, ‘ship length’ is the top RIF for the ‘minor’ consequence category, while ‘ship age’ is the top RIF for ‘catastrophic’ consequence. For all severity levels of consequences, it is considered possible to rank the most important RIFs as below:

Accident type > ship type > ship age > ship length > gross tonnage > visibility > traffic density > wind condition > ship flag > time of day > ship speed > equipment > lookout.

Furthermore, the real case scenarios can help examine the rationality of the developed model. We downloaded a reported case [56] from the ATSB website which was not included in the database described in Section 3.2. The information extracted from the report is as follows:

Occurrence time: shortly after 04:38 (local time); accident type: collision; ship type: bulk carrier; location: off Port Adelaide, South Australia; wind condition: light wind; visibility: good; lookout: bad; ship length: 108 m; gross tonnage: 6310; ship age: 38.

Figure 5 shows the results after the states of RIFs are updated to the trained model. The model gives a prediction of 88.4 percent of significant consequence, which is in line with the reality indicated in the investigation report.

4.3. Analytical Implications

The scenario analysis of this trained model can provide valuable predictions, especially concerning the most probable scenario. Figure 6 shows the outcome from the model. The ‘100%’ states in each node indicate the highest probability of occurrence. The figures in other states in each node indicate less chance of an incident occurring. Figure 6 shows that the most probable consequence is ‘minor’. And its preconditions are listed as follows:

(1)

Accident type ‘accident to person’ and ship type ‘general cargo ship’, ship length ‘less than 100 meters’, ship flag ‘non-FOC’, ship age ‘over 20 years’, gross tonnage ‘less than 3000’.

(2)

Time of the day ‘night’, ‘light’ wind, and visibility ‘good’.

(3)

‘Proper’ lookout, ‘safe’ speed, in ‘low’ traffic, with ‘good’ equipment.

The conditions above reveal that ship-related factors (e.g., ship age, ship length, ship speed) are highly associated with the accident consequence. Aged ships have aged equipment and an aged ship shell, and thus have a higher chance of becoming involved in an accident. Furthermore,

Table 8. Most probable scenarios.

RIFs	Minor	Significant	Severe	Catastrophic
Accident type	accident to person	others	accident to person	others
Ship length	<100	<100	<100	<100
Visibility	good	good	good	good
Wind condition	<5	<5	<5	>5
Ship age	>20	[11 15]	>20	>20
Ship flag	non-FOC	non-FOC	non-FOC	non-FOC
Gross tonnage	<3000	<3000	<3000	<3000
Traffic density	low	low	low	low
Ship type	general cargo ship	others	others	others
Time of day	night	day	day	day
Ship speed	safe speed	safe speed	safe speed	safe speed
Equipment	good	bad	good	bad
Lookout	proper	improper	improper	proper

outlines the conditions of all the most probable scenarios. Significant consequence tends to exist when ‘grounding’ accident occurs, particularly for the ‘others’ ship category with a length of less than 100 m, and ‘good’ lookout during the night time. In such situations, the vessel is probably navigating with a discursive bridge team. Certain high-risk conditions (e.g., off course, shallow water ahead) are probably overlooked by the duty navigators. When accidents are unavoidable, the navigators can do very little except undertake some mitigation measures. Grounding accidents tend to breach the ship hull and thus result in significant consequences. The accident reports ([57,58]) serve as illustrative examples of such cases.

Severe consequences also tend to occur for the ‘others’ ship type when ‘accident to person’ accident occurs in ‘day’ time, and the ship length is less than 100 m. In such conditions, some dangerous work is probably in progress on the vessel. This kind of work should comply with the safety management system on board, and thus, dangerous work permits are required. However, safe working practices are often not properly followed by duty teams ([59,60]) because they feel confident in their work. As a result, the working procedures, the supervision and the safety culture should be enhanced throughout the operation process.

Catastrophic consequences are most associated with the ‘others’ accident category. The accident report [61] serves as a typical case. Generally, a catastrophic consequence does not frequently occur; however, it can be disastrous when it does. ‘General cargo ship’ is the most likely involved ship type, and the ‘improper’ lookout and the ‘bad’ equipment condition are the key factors that should be properly stressed. These conditions must be well addressed when upgrading the safety management system.

Furthermore, the proposed model can serve as a predictor when facing threats. Users can adjust the state of the nodes to their conditions, generating the predicted types and consequences of accidents. By exploring various scenarios, analysts can identify high-risk situations and prioritize risk management strategies accordingly. It also can provide decision support by evaluating the consequences of different decisions or actions under various scenarios. This helps decision-makers to make informed choices based on the likely outcomes. By exploring different future scenarios, Bayesian network models can support strategic planning by identifying potential opportunities and threats.

In addition, the model can aid in resource allocation by identifying areas of the system that are most vulnerable or critical under different scenarios. This allows for a more effective allocation of resources to mitigate risks and improve system resilience. It also can be used to evaluate the effectiveness of different policies or interventions in achieving desired outcomes. By simulating various policy scenarios, analysts can assess their potential impacts and refine policies accordingly.

4.4. Comparative Analysis

The BN method has been utilized in many studies. The literature review in Section 2 shows that several other scholars have conducted similar research. Therefore, carrying out this comparative analysis to highlight the contributions of this study is essential.

Table 9. Comparison of studies.

Citation	Critical RIFs	Results and Findings
Li et al. [19]	Ship type; Ship operation; Voyage segment; Deadweight; Length.	Deadweight and hull construction are highlighted; Focus on ship accidents; Tankers tend to have ’collision’ accidents; Applied in global waters.
Sevgili et al. [18]	Accident type; Vessel age; Accident location; Distance to shore; Types of waterway.	Accident type and vessel age are highlighted; Focus on oil spill; Flooding and sinking tend to cause oil spills; Applied in US coastal waters.
Cao et al. [41]	Accident type; Ship type; Engine power; Gross tonnage; Location.	Accident type and ship type are highlighted; Focus on ship accident severity; Very serious accidents tend to be hull/machinery damage; Applied in global waters.
This research	Accident type; Ship type; Ship age; Ship length; Gross tonnage.	Accident type and ship type are highlighted; Focus on ship accident consequence; Severe accidents tend to be ‘accident to person’; Applied in global waters.

indicates the details of the comparison. The most critical RIF identified in [19] was ‘ship type’, while the other three studies placed ‘accident type’ in the first position. The five critical RIFs identified in this study also appear in the corresponding top five rankings of the other three studies, which implies the rationality of this research.

The comparison of the results and findings also show some interesting similarities and differences. In [19], the authors highlight the ‘deadweight’ and ‘hull construction’, but [41] and the present research emphasize ‘accident type’ and ‘ship type’. In contrast to [19], which analyzes ship accidents, paper [18] specifically focuses on oil spill consequences, while [41] and this research study examine the consequences of all ship accidents. Apart from the above, [19,41] apply to unlimited waters; however, the other two studies only apply to regional waters.

5. Conclusions

This study presents a maritime accident consequence analysis to investigate the influential mechanisms considering RIFs and consequence severity. The RIFs most relevant to consequence severity were identified based on the analysis of proposed model. The objectivity of this approach is considered a significant advantage. BN modeling is common in risk analysis across various fields, e.g., maritime safety, port security, freight supply chain, and climate change; however, its application in predicting consequence severity is relatively uncommon. As a result, the model is validated with various methods, including the Kappa test, comparative analysis, and scenario calculation.

The consequence severity, a topic that is rarely addressed in this field, is discussed in this paper. The critical RIFs are ranked: accident type, ship type, ship age, ship length, gross tonnage, visibility, traffic density, wind condition, ship flag, time of day, ship speed, equipment, and lookout.

However, this research does have certain limitations. The size of the database and the quantity RIFs are limited, which may restrict the model’s applicability in scenarios where the selected RIFs are not present. To enhance the flexibility and reliability of the research results, it is necessary to collect more data and include additional RIFs to support the development of the model. Furthermore, ergonomic factors, promotion pressures, etc., seldom appear in accident reports and thus are not included in this research. These factors will be the focus of future research.