Adaptation of Time-Triggered Systems using Neural Networks for Schedule Inference

Title: Adaptation of Time-Triggered Systems using Neural Networks for Schedule Inference

Abstract:
In recent years the field of safety-critical embedded systems has evolved towards novel application areas that combine stringent real-time constraints, reliability requirements and the need for runtime adaptation. Adaptation is a key factor for energy management, fault tolerance using fault recovery and service degradation. Adaptation also enables the dynamic incorporation of new components into the system based on an open-world assumption. At the same time, reliable operation and support for stringent real-time requirements are essential to support closed-loop control and guaranteed response times. In conventional safety-critical systems, time-triggered architectures provide significant advantages for developing dependable systems. The coordination of message exchanges is based on a communication schedule and a global time base in protocols such as Time Sensitive Networking and Time-Triggered Ethernet, thereby improving the temporal predictability, fault containment and safety certification. However, static schedules are not suitable for constructing adaptive systems that must react to context events.
Adaptive time-triggered architectures offer a solution to overcome these limitations by establishing at runtime suitable time-triggered communication plans, thereby combining the benefits of predictability, fault containment and safety with the support for context-dependent adaptation. We present the Adaptive Time-Triggered Multi-Core Architecture (ATMA), which supports adaptation using multi-schedule graphs while preserving the key properties of time-triggered systems including implicit synchronization, temporal predictability and avoidance of resource conflicts. We highlight the overall architecture for safety-critical systems based on a time-triggered network-on-a-chip and time-triggered off-chip networks with building blocks for context agreement, adaptation and deployment. Artificial neural networks are trained with feasible schedules for different types of context events, thereby enabling at run-time the inference on schedule changes. The architecture is evaluated using experimental scenarios and discussed for handling the requirements of warehouse shuttle systems and medical applications.