Safety Argumentation for ML-Enabled Perception Systems for Autonomous Trains State of the Discussion and Perspectives

Railway is constantly gaining importance as a sustainable mode of transportation.Highly automated train operation is a means to increase the utilization of existing networks.Technical solutions for driverless operation (Grade of Automation GoA3 and higher) typically rely on Machine Learning (ML) components to evaluate sensor data and establish situational awareness.Due to the inherent complexity and black-box nature of ML components, traditional approaches for safety argumentation are not directly applicable to ML-enabled perception systems.In our paper we present the state of the discussion on this subject and sketch-out potential approaches for technical solutions and the associated safety argumentation.First we discuss safety goals and objectives of perception systems for autonomous trains.We then highlight problem areas that prevent the use of traditional methods for arguing the safety of ML components and propose possible technical solutions and methods at ML component level and at the level of ML-enabled systems as a whole, taking into account the status of work in ongoing major funded projects.Finally we discuss strategies for safety argumentation and the role of safety argumentation as a driver for development decisions.