The market for autonomous maritime systems is growing steadily. Highly automated watercraft are taking on tasks such as cargo transport, coastal surveillance, marine research, environmental monitoring, and port logistics. These systems use artificial intelligence for object recognition and navigation. However, AI components can make unpredictable decisions or misclassify critical objects. Existing safety standards do not adequately address these challenges. Practical methods for risk analysis of AI in highly automated maritime systems are largely lacking.
In SIMAS, we are developing a methodology that enables a practical risk analysis of AI systems while taking into account the specific requirements of both the maritime industry and AI itself.
Bayesian Networks and FMEA for Analyzing AI Misclassifications
In the project “Development of a Methodology for the Analysis and Assessment of the Safety of Maritime Autonomous Systems (SIMAS),” we are developing a new methodology that combines classical safety analyses with probabilistic modeling. The approach is based on the direct integration of AI performance data—specifically, confusion matrices from object classification—into the probability tables of Bayesian networks. These are systematically linked to a Failure Mode and Effects Analysis (FMEA). This makes it possible to trace how misclassifications by an AI system propagate through the system architecture and what effects they have on overall safety.
The methodology also takes into account external factors such as weather conditions, sea state, or optical disturbances and assigns misclassifications to severity categories according to MIL-STD-882E.
Transparent Risk Quantification for Digitalization
This approach enables “what-if” analyses and design optimizations as early as the initial development phases. Developers can simulate various scenarios, identify vulnerabilities, and immediately quantify the impact of improvements—such as an increased detection rate. The results contribute to the safety case and help increase the acceptance of autonomous systems among regulatory authorities and classification societies.
Traceable, Quantitative Safety Assessment
The methodology offers developers and operators of maritime autonomous systems a decisive advantage: it enables a traceable, quantitative safety assessment that can be carried out with reasonable effort even as system complexity increases. Industry customers benefit from shorter approval cycles, reduced development risks, and a robust basis for justification when dealing with regulatory authorities. The approach is transferable to other AI components, such as navigation and route planning, and thus addresses a growing market in the maritime industry, the offshore energy sector, and beyond.
Simulations and experimental tests are used to validate practical applicability and lay the foundation for a robust safety assessment of AI-based maritime systems.
