A Deep Reinforcement Learning Decision Architecture for Predictive Maintenance in Smart Manufacturing

The continuous availability of critical assets is a fundamental requirement in modern manufacturing systems and Industry 4.0. Although recent advancements in deep learning have significantly improved the accuracy of equipment health prediction, a persistent challenge remains in translating these predictions into timely and cost-effective maintenance actions. This work addresses this prediction–action gap by reformulating maintenance decision-making as a sequential decision problem and introducing a Deep Reinforcement Learning (DRL)–based decision architecture for predictive maintenance. Equipment health dynamics are modeled as a Markov Decision Process (MDP), and a Deep Q-Network (DQN) agent is trained to learn maintenance policies that balance failure avoidance and production continuity over time. The proposed framework is validated on a high-dimensional public industrial pump dataset, where the learned policy achieves a recall of 96.6% under noisy operating conditions. The results show that policy-based DRL approach can be used complementary to conventional threshold-driven maintenance strategies, highlighting their potential as a decision-support layer in real-world predictive maintenance systems.