A Condition-Aware Scheduling Framework for Flexible Manufacturing Systems Using Deep Reinforcement Learning

Abstract

Factory condition-monitoring data are often available to maintenance dashboards but absent from production scheduling decisions. This paper proposes the Condition-Aware Scheduling Framework (CASF), an architecture aligned with ISA-95 (the ANSI/ISA-95 enterprise-control system integration standard) that exposes machine health to scheduling without requiring a new reinforcement-learning algorithm or a new raw-sensor-based policy. CASF separates a Condition Abstraction Layer (CAL) from a Condition-Aware Scheduler (CAS) trained with deep reinforcement learning (DRL). The CAL converts sensor-derived degradation evidence into a discrete fleet-condition state. The CAS combines this state with shop-floor production features and an operational feasibility mask to choose dispatching and condition-based maintenance (CBM) actions. We instantiate the CAL as a root-mean-square (RMS) vibration extractor calibrated from XJTU-SY bearing data and train the CAS with synthetic degradation trajectories. Simulation experiments under in-distribution and scale-shifted settings show that condition-aware policies improve maintenance-sensitive scheduling decisions over state-threshold and condition-blind alternatives. Online inference remains compatible with Manufacturing Execution System (MES)- or edge-level deployment. Under a paired fixed-set protocol, condition-aware DRL significantly outperforms the state-threshold baseline on 40 of 60 (configuration, scale) cells. Withholding machine-health state degrades reward three- to sevenfold on in-distribution scales. The main contribution is architectural: a stable cross-layer condition abstraction connecting existing sensing infrastructure to production scheduling authority.