A methodology for optimizing Time-Sensitive Networking scheduling schemes based on a dual graph attention network

Time-Sensitive Networking (TSN) is pivotal for enabling deterministic communication in industrial automation and cyber-physical systems, where network component failures can cause severe service disruptions. However, due to the complex and dynamic nature of flows, designing optimal scheduling schemes that simultaneously satisfy constraints on end-to-end latency and reliability remains a challenge. To address this, this paper proposes a novel methodology for optimizing TSN scheduling based on a Dual Graph Attention Network (Dual-GAT). The Dual-GAT is incorporated to enhance feature extraction for scheduling and improve scheduling accuracy. The optimization of scheduling schemes is implemented in the TSN scheduling framework abstracted as a Markov decision process (MDP). Unlike traditional DRL approaches that struggle with spatial graph dependencies, our Dual-GAT acts as a feature encoder and cooperates with the MDP framework to jointly process flow requirements and hardware port constraints through parallel attention blocks. This architecture enables the agent to perceive intricate spatial relationships and structural topology metrics. This data-driven framework learns to generate scheduling strategies that inherently minimize latency and maximize reliability. As a case study, TSN scheduling of the in-vehicle centralized electrical/electronic architecture is used to demonstrate the applicability of the methodology. The proposed model is beneficial to optimize the design of the in-vehicle electrical/electronic architecture.

Keywords: In-vehicle network failure, Time-Sensitive Networking, Graph Attention Network, Deep reinforcement learning, Scheduling strategy optimization, rerouting mechanism.