Product Line Engineering Applied to Perception System Architectures for Autonomous Trains

Autonomous trains rely on on-board perception systems that allow to identify the train's location and route and that are able to detect potential obstacles on the track used by the train. To guarantee the required safety integrity levels for perception systems employing neural networks, N-version perception system architectures may be used that combine parallel processing pipelines, using design, implementation and/or data diversity to detect or even mask individual perception errors. In our paper we present an approach to systematically specify perception system architectures using feature-based product line engineering, and to link the feature and variant models to architectural models of perception systems, in order to automatically create perception system architecture instances. We model architectural features and the related restrictions to specify valid architectures, and define perception architecture variants. Via the connection between the product line engineering models and systems engineering specifications of the computational elements used in the architectures, we can transform the architecture variant specifications into systems engineering architecture models and can automatically generate the software source code and configuration files required for implementation of the architecture variants. Our approach supports the systematic exploration of potential perception system architecture variants and the efficient evaluation of their individual perception quality and robustness against certain classes of perception challenges.