Risk Assessment for Autonomous Ships Using an Integrated Machine Learning Approach ^†

Abstract

The inherent complexities of Artificial Intelligence (AI) and machine learning (ML) technologies expose autonomous ships to a wide range of multifaceted interconnected risks. However, very few studies have aimed at the holistic risk assessment of autonomous ships. To this end, this study employs an expert-opinion-based integrated machine learning approach amalgamating logistic regression and Bayesian network to conduct risk assessment for autonomous ships. The results reveal human factor interactions and operational issues as the prominent accident causation factors. The findings of this study will contribute significantly to the existing literature on autonomous ships and the complexities involved in their operational systems. By identifying critical factors causing accidents and their impact on autonomous ship safety and resilience, stakeholders such as autonomous ship manufacturers, port authorities, shipping companies, and governments can develop more efficient and effective operational and safety systems.

1. Introduction

Technological advancements in Artificial Intelligence (AI) and machine learning (ML) have played a crucial role in the development of autonomous ships [1]. As per the definition by maritime autonomous surface ships (MASSs) of the International Maritime Organization (IMO), the autonomous ships are equipped with a system with the capability to make decisions and contrive pertinent actions [2]. Sensor technology plays a vital role in the functioning of autonomous ships, as they heavily rely on a range of sensors such as radar, lidar, sonar, and cameras to perceive and interpret the surrounding environment [3,4]. However, such intricate systems are associated with pertinent sophisticated risks.

Evading the dependency on human factors will have a substantial impact on the safety augmentation, as 60% to 90% of the accidents in the maritime transportation are attributed to human factor [5,6]. Nevertheless, the human and autonomous ship technology interaction is interlinked with several complexities and risks [7,8,9]. The multifaceted functionality of autonomous ships introduces a host of intricate risk factors that can potentially lead to accidents [10,11].

Autonomous systems heavily rely on complex software algorithms and hardware components. Malfunctions, bugs, or errors in the software or hardware can lead to system failures, loss of control, or incorrect decision making, potentially resulting in accidents [12]. Autonomous systems are susceptible to cyber security threats, including hacking, unauthorized access, or malware attacks [13]. Breaches in system security can compromise the control, navigation, or communication of autonomous ships, leading to accidents or malicious activities [14,15]. Autonomous systems involve the integration of multiple components, including software, hardware, sensors, and communication systems [16]. The complexity of integrating these components introduces challenges in ensuring their seamless operation, compatibility, and reliability, which can contribute to accidents if not adequately addressed [17].

Autonomous systems make decisions based on algorithms and predefined rules. However, handling complex and unpredictable situations, uncertainty in sensor data, or encountering novel scenarios can pose challenges for autonomous decision making, potentially leading to accidents or suboptimal responses [18]. The legal and regulatory framework surrounding autonomous systems is still evolving. The absence of comprehensive regulations or standards specific to autonomous ships may create uncertainties and challenges in ensuring safe operations, liability assignment, and adherence to maritime rules [19].

In front of this backdrop, the current study is aimed at the risk assessment of autonomous ships, incorporating a holistic approach towards risk causation factors, which encompass technology failures, cyber security threats, inadequate regulations, environmental challenges, human error in design and maintenance, interactions with other vessels, and legal and ethical dilemmas. This study will incorporate an expert judgment- and literature-based integrated supervised machine learning approach amalgamating binary logistic regression and Bayesian networks.

This study will make significant contributions by examining key factors, and a thorough understanding of the risks and accident causation factors associated with autonomous ships can be achieved. These insights can inform the development of industry standards, regulations, and guidelines, facilitating the widespread adoption of autonomous ships while ensuring the highest level of safety and risk management. This study will support the development of a robust and reliable autonomous ship industry, reinforcing the importance of proactive risk management and effective accident prevention measures, guidelines, and regulations, thereby advancing the safety and reliability of autonomous ship operations.

2. Methodology

The complete process of modeling and evaluating is conducted within the Bayesian network interface software known as Hugin, which offers a robust and dependable environment for conducting probabilistic risk calculations and inference modeling. The comprehensive methodology is outlined in a step-by-step manner, as follows.

2.1. Variable Determination

The variables are determined subject to expert judgment, accident reports, and the literature. The relationships between variables are established in such a manner that the variables causing the effects are denoted as parent nodes, while the resulting effect variables are referred to as child nodes. Arrows originate from the parent nodes and terminate at the child nodes, illustrating the causal connections between them.

2.2. Joint and Conditional Probability Quantification

After establishing the graphical model, the next step involves developing the Conditional Probability Table (CPT) for each node. In this study, the methodology is grounded in binary logistic regression and expert judgment, as elucidated and detailed by one of the authors in a paper addressing the ship safety index [20].

2.3. Posterior Probability Determination

The critical probabilities within a Bayesian network can be derived using the following equations.

The distribution of joint probability can be found by

$$P\left(Y=y_q,X_p=x_{pq}\right)=P(X_p=x_{pq})\times P(Y=y_q|X_p=x_{pq})$$

(1)

whereas the marginalization is found by

$$P\left(Y=y_q\right)={\sum }_q^{n}P(X_p=x_{pq})\times P(Y=y_q|X_p=x_{pq})$$

(2)

The prevailing Bayesian rule is illustrated as

$$P\left(X_p=x_{pq}|Y=y_p\right)=\frac{P(X_p=x_{pq})\times P(Y=y_q|X_p=x_{pq})}{P(Y=y_q)}$$

(3)

The quantification of β values is achieved through binary logistics regression, whereas Equation (2) provides the conditional probability. By incorporating expert judgment to establish the factors and variables data, the values for Equation (3) can be computed accordingly.

3. Autonomous Ship Risk Assessment

The directed acyclic graph (DAG) graphical illustration of the developed model is illustrated in Figure 1, all the factors or variables are binary having states “0” and “1”, indicating “No” and “Yes”, respectively.

After developing the model in a Hugin environment, the next step is to put in the values of probabilities quantified through the expert judgment and binary logistic regression. Before running the model, it is first validated. The change brought in the probabilities of parent node induced a subsequent change in the probabilities of child node, and the influence of evidence was higher than the sub evidence. After validation, the model is run and the inference results without setting evidence are analyzed. As evident from Figure 2, the autonomous ship accident risk has a probability value of 6.77 where human factor, cyber security issues, and operational issues remain the highest accident causation factors with a contribution incidence value of 14.13, 12, and 11.44, respectively. Moreover, human–system interaction, decision-making uncertainty, and system complexity integration issues are critical causation sub factors, with incidence values of 35, 35, and 30, respectively.

Autonomous ships rely on a combination of human operators and automated systems. The interaction between humans and the autonomous system is critical for effective decision making and safe operation [21]. If there are flaws or misunderstandings in the communication between humans and the system, it can lead to errors, misinterpretations, and accidents. Clear protocols, interfaces, and well-defined roles are necessary to establish effective human–system interaction. Decision-making uncertainties can arise from unpredictable factors like weather, other vessels, or technical malfunctions. Flaws in decision-making algorithms may not account for all potential scenarios, leading to incorrect or inappropriate responses in critical situations.

Autonomous ships consist of numerous interconnected systems, including sensors, control systems, navigation systems, and communication networks. The complexity arises from the integration of these systems and their interactions. If there are design flaws, software bugs, hardware failures, or inadequate system integration, it can result in malfunctions or unexpected behaviors. These issues can compromise the safety of the autonomous ship, leading to accidents.

Now, using the inverse propagation feature of the Bayesian networks, evidence is set at the autonomous ship accident to occur by turning its incidence likelihood to 100. This aspect facilitates the determination of the highest-contributing and most critical accident causation factors in autonomous ship accident scenario. As is evident from Figure 3, if evidence is set at the autonomous ship accident to occur, the contribution likelihood of human factor remains the highest with a probability value of 23.61, whereas operational issues remain the second highest contributor in this scenario with an incidence value of 20.13. Likewise, the cyber security issues remain the third highest causation factor with incidence value of 16.12. However, if we take into account the direct inference values, the comparison with the inverse propagation value indicates that the operational issues underwent the highest increase in incidence values, increasing from 11.44 to 20.13, which is a 76% increase in its causation role.

Sensitivity Analysis

In this study, the sensitivity analysis is conducted by observing the percent change in the occurrence likelihood of the autonomous ship accident risk subject to the minimum and maximum role of every factor under consideration in the existing model and the most critical factors are reported in

Table 1. Sensitivity analysis values for the most critical accident causation factors.

Variable	Minimum	Maximum	Percent Change
Operational Issues	6.10	11.91	95
Human Factor	6.02	11.31	82
Cyber Security Issues	6.45	9.09	41
Technical Issues	6.57	8.67	32

. The tabulated results reveal that operational issues and the human factor are the most sensitive and critical factors for autonomous ship safety, with sensitivity values of 95 and 82, respectively. Likewise, the existence of cyber security threats and technical issues also needs to be given due attention. In order to ensure safer and resilient autonomous ship operations, the associated stakeholders, including autonomous ship designing and building companies, liner companies, port authorities and governments, need to focus on the critical factors and devise concurrent safety measures after conducting detailed investigations of each aspect and scenario.

4. Conclusions

This study employs an expert-judgment-based integrated supervised machine learning approach to conduct a holistic risk assessment for autonomous ship operations. The prominent results of the study are concluded as follows.

• Without setting evidence, the autonomous ships are exposed to an accident risk probability of 6.77, where human factor, operational issues, and cyber security issues remain the highest accident causation factors.
• The inverse propagation of the Bayesian network indicates that for an autonomous ship accident to occur, human factor and operational issues remain the highest contributors, with the role of operational issues undergoing highest change in incidence.
• A sensitivity analysis was conducted to reveal the most critical and sensitive factors for the safety and resilience of autonomous ships.

The study acknowledges its limitations and suggests future research directions, such as exploring additional factors and incorporating risk categorization and damage level classification.