FusionNet Lite: Compact Federated Deep Learning for Predictive Maintenance under Label Heterogeneity and Temporal Generalization

Abstract Federated learning (FL) supports collaborative predictive maintenance (PdM), but compactness alone does not establish reliable failure detection or consistent explanations. We evaluate FusionNet~Lite, a 1,420-parameter feature-axis classifier, against four lightweight baselines under centralized training and Federated Averaging (FedAvg). Ten matched core seeds and validation-selected checkpoints are used with stratified AI4I~2020 partitions and a fixed-row chronological MetroPT3 holdout. Separate studies extend the three-client core to ten clients and examine architectural, optimization and participation sensitivities. On MetroPT3, FedAvg FusionNet~Lite achieves area under the receiver operating characteristic curve (AUROC) $0.9351\pm0.0462$ and area under the precision--recall curve (AUPRC) $0.2197\pm0.1971$, expressed as mean and sample standard deviation. Switching from original-grid validation-F$_1$ to expanded-grid validation-F$_2$ selection changes mean test F$_2$ from $0.0423\pm0.0995$ to $0.1566\pm0.2044$, retaining substantial variability. FusionNet~Lite has the smallest parameter count and lowest fixed-batch parameter--latency index among the five core architectures, while predictive rankings depend on dataset and configuration. Paired, class-specific Integrated Gradients (IG) comparisons across all five models show variable central--federated agreement. These findings support a conditional performance--complexity trade-off within a controlled benchmark. Forecasting lead time, physical energy savings and industrial deployment readiness remain unestablished; future work requires independent equipment, alternative aggregation and device-level validation.