Project duration: 01.01.2022 - 31.12.2024 | Website: safetrain-projekt.de

For better quality and scope of train services, higher levels of automation need to be achieved in train operations. Although trains can already operate partially autonomously in controlled environments such as airports, this is not yet happening in the free world outside controlled zones. safe.trAIn is seeking to develop solutions to this complex challenge in order to harness the potential of automated train operations for safety, reliability and sustainability. Artificial intelligence is an important component of future control platforms for driverless train operation, such as in obstacle detection.

To ensure the usability of AI solutions on trains, the safety and transparency of the underlying algorithms must be demonstrated. The safe.trAIn consortium, which consists of various industry partners, software companies and regulatory authorities, is addressing these issues in order to ensure high-quality train transportation. Therefore, safe.trAIn also accelerates the transition to an environmentally friendly transportation system. 

In safe.trAIn, Merantix Momentum is developing methods for the safe AI function of autonomous train control. In particular, the focus is on label efficiency and robustness as well as transparency of the algorithms. Due to the scarcity of annotated data in the train domain, methods need to handle labels as efficiently as possible. Merantix Momentum develops approaches to tailor neural networks that have been pre-trained with large data sets to the project-specific data sets.

The recently popular methods of self-directed learning are particularly interesting for this purpose, since they do not require explicit labels for fine-tuning. Robustness and transparency are particularly important for the project, as without them, safe train operation is not possible. Merantix Momentum is investigating the impact of out-of-distribution data on downstream models and how such inputs can be detected. Furthermore, the project aims to create more favorable properties with respect to transparency through special network architectures.