Quantitative Risk Assessment for Maritime Autonomous Navigation using AI Perception Metrics and Bayesian Networks

As maritime navigation increasingly shifts towards autonomy, ensuring the safety of AI-enabled perception systems under complex and uncertain conditions is a critical challenge. Traditional safety analysis methods often fail to capture the variability and uncertainty introduced by AI components in maritime environments. In this study, we investigate whether system-level risk for autonomous maritime navigation can be quantified by embedding objectclassification performance metrics into a Failure Mode and Effects Analysis (FMEA)-linked Bayesian Network (BN). This network allows us to estimate the probability of failures, and to describe risk using these calculated probabilities together with a scenario-based model of severity. The novel framework integrates empirically derived AI performance data, specifically confusion-matrix probabilities from a computer-vision model, directly into the conditional probability tables of a Bayesian Network for Maritime Autonomous Surface Ships (MASS). In addition to internal AI components, the framework also models external factors, such as adverse weather conditions, and evaluates the impact of misclassifications on navigational safety. Initial results confirm the feasibility of the integrated framework and its capability to propagate AI misclassification probabilities through the Bayesian Network to derive a system-level risk consistent with MIL-STD-882E. By parameterizing the network using confusion-matrix information and reusing Bayesian structures across scenarios, the approach provides, for the first time in a maritime context, a transparent and scalable basis for risk-based assurance of AI-enabled maritime systems. Our main goal is to provide developers with confidence that architectural and design choices lead to systems that are free from unacceptable risks by quantifying uncertainties and allowing for uncertainty-aware safety claims, thereby yielding evidence that these claims hold. The proposed method bridges qualitative and quantitative safety perspectives by combining structured FMEA reasoning with Bayesian probabilistic inference. In future work, we will validate the approach in autonomous maritime trials to establish a robust foundation for assessing AI-enabled navigation safety.

Keywords: Probabilistic risk assessment, AI perception systems, Maritime autonomous navigation, Maritime Autonomous Surface Ships (MASS), Bayesian Networks.