Real-Time Safe Production Scheduling via Causal Inference and Hybrid-Event Control

This paper proposes a real-time and certifiably safe scheduling framework for cyber-physical manufacturing systems, designed to operate under hybrid discrete–continuous dynamics and deployment-level latency constraints. The approach integrates two core components: (i) awavelet-invariant causal modeling modulethat uncovers multiscale dependencies in sensor data while preserving conditional-independence structure under orthonormal wavelet projections, and (ii) ahybrid-event barrier certificate (HRCBC) filterthat enforces hard safety constraints across both continuous flows and discrete event triggers such as tool changeovers and emergency halts. Departing from traditional methods that assume smooth dynamics or fixed planning horizons, the architecture fuses offline causal discovery with fast, online, safety-filtered control. Causal graphs are learned via orthonormal wavelet projections and refreshed per shift or upon concept-drift detection. Real-time safety is maintained by solving a convex quadratic program with warm starts and p99 latency under 10 ms. The method is validated on two industrial lines (Site-A: EV-module assembly; Site-B: appliance assembly) and on public synthetic benchmarks (SynTwin-Bench), with ablations and multi-site simulations. Across the reported testbeds, WaIST+HRCBC improves cycle time by up to 28% and operating cost by 20% relative to the incumbent dispatcher; throughout the reported evaluation horizon, no safety violations were recorded by the independent PLC safety layer, while maintaining PLC-compatible latency. The system is deployable on standard industrial gateways and interoperable with MES/PLC via RESTful and OPC UA adapters.