FL‐Offshore: A Federated Learning Framework for Predictive Maintenance in Brazilian Offshore Production Scenario

ABSTRACT Offshore oil production in Brazil relies on Floating Production Storage and Offloading (FPSO) units, where failures in rotating machinery motivate machine learning–based predictive maintenance. Centralised training is impractical due to privacy constraints, heterogeneous and highly imbalanced local datasets and unreliable satellite connectivity. This work investigates federated learning for predictive maintenance in Brazilian offshore FPSO environments using a full factorial experimental design that models data heterogeneity, class imbalance and intermittent client participation. Adaptive federated optimization strategies outperform FedAvg under unreliable connectivity, with FedAdam achieving the best overall performance and an average improvement of 8.5% over FedAdagrad, whereas FedAdagrad preserves higher recall in severely imbalanced scenarios. The results demonstrate that the FL‐Offshore framework is a realistic and effective platform for evaluating federated learning under offshore operational constraints, enabling stable and accurate fault detection with limited connectivity.