From Prediction to Decision: Dynamic Maintenance Scheduling in Industrial Edge Computing via Federated DRL

Abstract

In industrial edge computing, numerous devices must operate continuously to support time-critical services. Their degradation can lead to localized failures and reduce overall system uptime, making proactive maintenance necessary. Existing data-driven maintenance methods typically predict device health or remaining useful life using deep learning models and then trigger maintenance through predefined thresholds and heuristic rules. However, this open-loop paradigm cannot dynamically adapt its decisions to evolving conditions and struggles to scale maintenance intelligence across distributed edge sites, especially under data privacy constraints. This paper transitions from prediction to decision by reformulating maintenance scheduling as a closed-loop sequential decision-making problem. Specifically, we formulate a Long-Term Operational Cost Minimization (LTOCM) problem over three device operational modes to directly capture the trade-off between failure loss and maintenance cost. Based on LTOCM, we construct a Markov decision process with a cost-based reward and further introduce a degradation-aware reward to provide informative learning signals for maintenance timing. We then propose FDDM, a federated deep reinforcement learning scheme in which edge agents learn closed-loop maintenance policies locally and collaboratively improve them across distributed industrial edge sites without sharing raw operational data. Momentum-based global aggregation is further incorporated to stabilize federated policy learning. Extensive experiments on the NASA C-MAPSS dataset across multiple fault scenarios show that FDDM achieves near-optimal operational cost and faster convergence than representative baselines. Ablation and non-IID evaluations validate the effectiveness and robustness of the proposed design.