Fraunhofer IKS is focusing on three particular topics in the safe.trAIn project:
- Proof of safety for AI functions in the perception of the train
- Robustness of AI and monitoring mechanisms for AI-based systems
- Operational design domain (ODD) specification and monitoring for rail transport
In this context, Fraunhofer is involved in two work packages in particular:
Proof of safety for AI methods
Within the safe.trAIn project, one of the responsibilities of Fraunhofer IKS include the test methods and tools that prove the reliability of AI methods. Fraunhofer IKS is developing methods and tools based on the properties and acceptance criteria for AI systems specified in the project to demonstrate the explainability, robustness and compliance with safety requirements. Special metrics and concrete methods are being researched, such as the reliable detection of humans or the detection of unknown situations (out-of-distribution detection).
Fraunhofer IKS is also working on methods that assure the quality of AI functions as early as during the development phase. A key point of research is on methods that increase the transparency and interpretability of AI applications.
Safety architecture for AI-based functions
Fraunhofer IKS is researching the safety architecture for AI-based functions in fully automated operation (GoA4). In order to do this, the requirements for the driverless regional train are first defined and then a Technical Safety Plan (TESIP) is created. The operational design domain (ODD) must also be determined, which records the environmental conditions, including individuals and other systems interacting with the system, as well as the operational, climate and weather conditions. The ODD specification is pivotal for the driverless regional train’s safety justification and vehicle architecture. Fraunhofer IKS is concentrating its efforts in this area on researching a method for specifying and monitoring the ODD for rail transport. The ultimate goal is a safe, reliable, available, and efficiently maintainable overall system.