Threshold-grounded decision reliability for predictive maintenance under centralized and federated learning

Abstract

This capsule supports independent verification of a threshold-grounded predictive-maintenance (PdM) reliability study. The work asks whether maintenance decisions remain defensible under a fixed operating threshold when industrial information is uncertain, class-imbalanced, and split across sites, comparing centralized training with Federated Averaging as a deployment constraint rather than as a new aggregation method. The hosted Reproducible Run executes smoke mode: tiny synthetic caches and a reduced model subset confirm that the training, metric, and registry pipeline runs on a clean machine. Smoke output is not paper evidence. The authors' completed 126-run industrial registry is archived at expected_outputs/reference_registry/master_results_rerun.csv. Full regeneration requires prepared leakage-safe caches under data/artifacts_cache/ as documented in data/raw_placeholders/README_DATASETS.md. Dependencies install from environment/requirements.txt at run time (Python 3.10, PyTorch, scikit-learn, pandas, and supporting libraries). Do not place zip archives inside /environment. The driver script is code/run, which calls python -u scripts/run_reproduce.py --mode smoke --no-skip-existing. Revisions Added: One new file was added. It does not replace the frozen trainers: code/jp4_revision_r1.py Unchanged from the published capsule. Central and FedAvg training, the smoke Reproducible Run, the seed-42 reference registry, the IMS label recipe, and independent per-dataset evaluation.