Fault-Tolerant Real-Time Scheduling for Edge AI in US Critical Infrastructure

Abstract

The integration of Edge AI promises to enhance operational efficiency within US critical infrastructure, yet it simultaneously introduces significant challenges regarding reliability and real-time determinism. While existing scheduling methods are suitable for general IoT environments, they frequently fail to distinguish between safety-critical and non-critical tasks, potentially resulting in severe consequences following computational failures. To address this challenge, this paper proposes a framework termed Safety-Critical Graph Reinforcement Learning, specifically designed to handle fault-tolerant real-time scheduling in highly dynamic and adversarial edge environments. Unlike traditional responsive methods that rely solely on task migration, SC-GRL integrates an active primary-backup mechanism and permissible model degradation into the action space of a Proximal Policy Optimization agent. By modeling the edge topology as a dynamic graph with embedded criticality attributes, the SC-GRL agent learns to optimize a composite objective that prioritizes deadline satisfaction for high-criticality tasks over mere resource utilization. Extensive simulations utilizing real-world trajectories from the US energy and transportation sectors demonstrate that SC-GRL significantly reduces the deadline miss rate for critical tasks under heavy loads and random node failures compared to state-of-the-art graph-based baselines. The data points to a clear conclusion here. If we want to build truly resilient public infrastructure, any deep reinforcement learning we use for scheduling must have a built-in sense of what's most critical.