Scalable decentralized prognostics for industrial systems under data heterogeneity

Condition-Based Monitoring (CBM) plays a vital role in predictive maintenance by enabling early fault detection through real-time sensor data analysis. However, the rarity of fault events in industrial systems limits the performance of centralized learning approaches, which often overfit to normal conditions and miss rare failures. Centralized methods also raise privacy, communication, and scalability concerns. The convergence of global models in federated settings is influenced by the distribution of fault data across local devices. In practical deployments, this distribution is often non-uniform, which can hinder convergence. To address these challenges, this study introduces a federated learning (FL) benchmark tailored for condition-based monitoring of sleeve bearings under realistic data-scarce fault scenarios. Rather than relying on conventional independent and identically distributed (IID) assumptions, we design controlled non-IID data distributions using Dirichlet sampling applied to real sensor datasets. This enables systematic exploration of how varying degrees of heterogeneity influence FL performance. We benchmark multiple base, scaled-up, and novel aggregation strategies across deep network architectures, capturing both classification and remaining useful life prediction tasks. Crucially, we expose how the Dirichlet α parameter interacts with optimizer-specific dynamics, revealing failure modes under moderate non-IID conditions and identifying regimes where FL remains stable or collapses. By bridging empirical evaluation with deployment-relevant scenarios, our study provides actionable heuristics for FL-based CBM in resource-constrained, privacy-sensitive industrial environments. • Deployment-driven evaluation framework for federated learning in condition-based monitoring (FL-CBM). • Systematic analysis of aggregation strategies under non-IID and fault-scarce conditions. • Dirichlet-based data partitioning for controlled heterogeneity and optimizer stress testing. • Unified evaluation of fault classification and RUL prediction under client heterogeneity. • Empirical thresholds and design guidelines for stable FL deployment in industrial environments.