Semantic-Aware Time-Sensitive Networking Scheduling with Constrained Safe Reinforcement Learning

Time-Sensitive Networking (TSN) provides deterministic communication guarantees for industrial automation through mechanisms such as IEEE 802.1Qbv time-aware shaping and frame preemption. However, existing TSN scheduling methods optimize exclusively for network-layer metrics—throughput,jitter, end-to-end latency—ignoring the task-level semantics of industrial agents. In a multi-agent factory, the same sensor stream may tolerate 10 ms latency during routine inspection but require sub-millisecond delivery when gating an emergency stop. This semantic gap causes a fundamental misalignment between network resource allocation and application-criticality. We present the first semantic-aware TSN scheduling framework that bridges this gap through three integrated components. First, a task intent ontology maps six industrial task types (inspection, collaboration, emergency stop, periodic control, telemetry, reconfiguration) across four criticality levels (L0–L3) toflow semantic descriptors—priority weights, urgency functions, and delayable boundaries—which are then translated to GCL and CBS parameters. Second, a Constrained Safe Reinforcement Learning (CSRL) scheduler, formulated as a Lagrangianconstrained MDP, optimizes a semantic-weighted task completion rate while treating Network Calculus (NC) worst-case delay bounds as a learnable constraint. Third, an NC safety shield validates every scheduling action against deterministic schedulability constraints before execution, providing policy-independent safety with per-criticality tolerance policies. Evaluated on a 3-switch line topology with a 10 ms hyperperiod under three experiment modes—static benchmark (5/8/12 flows), dynamic flow arrival (5 base + 3 arriving), and component ablation—the CSRL scheduler achieves 100% task completion at 5 and 8 flows and 97.5% at 12 flows (92.8% ± 3.6% across three seeds), close to a static greedy GCL baseline that achieves 100% in all static scenarios. The Lagrangian multiplier λ, which couples the Network Calculus constraint into the learning objective, rises monotonically with flow density (0.12 at 5 flows to 0.67 at 12 flows), demonstrating that the formal constraint pressure is genuinely transmitted into the learned policy. Ablation studies confirm the Lagrangian mechanism is active (Full CSRL λ = 0.29 vs. constraint-disabled λ ≡ 0) and that semantic weighting activates the constraint term more strongly (No-Semantic λ = 0.10). These results establish a semantically-aware, formally-verifiable RL scheduling architecture whose performance is competitive with classical schedulers while retaining online adaptability