A holistic approach to risk assessment for maritime autonomous ship operations

Abstract

Maritime Autonomous Surface Ships (MASS) are expected to transform maritime transport by reducing crew workload, enabling new business models and potentially improving safety performance. At the same time, the transfer of navigational tasks from bridge teams to automated and remotely supervised systems introduces new types of hazards and reshapes traditional ones. Decision-makers therefore need risk assessment tools that can cope with sparse data, heterogeneous evidence and rapidly evolving technology, while remaining transparent enough to support design, regulatory and operational decisions for early MASS deployments. This thesis develops and applies a holistic framework for quantitative risk assessment of MASS navigation. The work concentrates on ship-level navigational risks that may lead to collisions, groundings or loss of control during open-sea and coastal transits, including operations in mixed traffic with conventional vessels. The focus is on the core navigational functions of sensing, decision-making and control, and on the ways in which different degrees of autonomy affect the risk picture. Four broad failure categories are defined: Human-Related, Technological, Operational and Environmental conditions. The starting point is a structured taxonomy of 16 navigational failure modes derived from a PRISMA guided systematic review of autonomous-ship safety studies, regulatory and technical documents, and expert input. Human-related modes include insufficient navigational skills, resistance to new technologies, cognitive overload and inadequate training on digital tools. Technological failures cover sensor malfunctions, navigation software glitches, cybersecurity breaches and connectivity issues. Operational failures comprise over-reliance on automation, inadequate emergency response, inefficient route planning and failure to detect small obstacles. Environmental conditions include adverse weather, strong currents, sea-temperature fluctuations and restricted visibility. These failure modes are assessed through a Failure Modes, Effects and Criticality Analysis (FMECA) that uses three risk parameters: Occurrence, Severity and Detectability. Ratings are expressed on five level linguistic scales rather than as point values, to reflect the lack of historical data for MASS operations. Judgements are elicited from a purposively sampled panel of seven domain experts with experience in ship navigation, remote-operation concepts, safety management and regulation. Individual expert opinions are then translated into basic belief assignments in the sense of Dempster Shafer (DS) evidence theory, providing an explicit representation of both support and ignorance. In the second step, DS fusion rules are used to aggregate the experts' belief structures for each failure mode. The combined mass functions are subsequently mapped to a dimensionless Crisp Risk Value (CRV) index on a 0–125 scale, which condenses the joint effect of Occurrence, Severity and Detectability while preserving the relative ordering of hazards. The resulting ranking highlights a small subset of highly critical failure modes. Sensor malfunctions obtain the highest CRV, around 45, closely followed by over-reliance on automation and cybersecurity breaches with CRVs in the low-40s. At the category level, Technological failures emerge as the most critical group with a CRV close to 35, followed by Human-Related and Operational failures with CRVs around 30 and Environmental conditions with CRVs in the low-20s. The third step links the DS results to a rule-based Bayesian Network (BN) that models causal relationships between the four failure categories and an overall MASS navigational-risk node. A transparent set of IF–THEN rules translates combinations of Occurrence, Severity and Detectability ratings into discrete risk states that serve as inputs to the BN. Once the network is parametrised with CRVs and prior probabilities, it can be used to update the risk picture under alternative scenarios, propagate uncertainty and explore how changes in specific failure modes or categories influence the overall navigational-risk level. Quantitative analyses performed on this model show that the baseline MASS navigational risk is driven primarily by the interaction of technological and humanrelated failures, with environmental conditions acting mainly as amplifiers of existing vulnerabilities. Sensitivity factors computed at the category level indicate that modest improvements in technological and human-related failures would yield the largest reductions in overall CRV, whereas changes in environmental conditions have a smaller but still noticeable effect. Structural robustness tests, including a scenario in which over-reliance on automation is reclassified between categories, confirm that while category-level CRVs shift, the identity and relative ordering of the most critical failure modes remain stable. Overall, the thesis demonstrates that an integrated FMECA–DS–rule-based BN pipeline can deliver a coherent quantitative picture of navigational risk for MASS even when empirical data are scarce. The framework offers a structured way to prioritise hazards, clarify the impact of expert disagreement and uncertainty, and test the stability of conclusions under plausible changes in assumptions. The findings point to sensor reliability, cybersecurity protection, thoughtful automation design and training for digital navigation tools as key levers for improving the safety of autonomous and highly automated ship operations. The approach is designed to be reusable and updatable as new operational data and experience with MASS become available, providing a basis for evidence-informed design choices, regulatory development and operational guidelines.