Safe Usage of Multi-Cores in Neural Network Avionics Applications

While the European Union Aviation Safety Agency pushes towards artificial intelligence and neural networks, their introduction is still prevented by a lack of reliable high-performance platforms. This paper addresses this gap by adopting the novel Timing Diversity approach from the real-time community and adapting it to the avionics domain to achieve a reliable and safe timing behavior on system-level. Timing Diversity provides a safety net to handle unexpected timing outliers of multi-core processors by exploiting existing modular redundancy. The system design is illustrated using a neural network-based runway detection, deploying different multi-core hardware and diverse software algorithms. Furthermore, error handling and compliance to current acceptable means according to AMC 20-193 are discussed.