Predictive maintenance (PdM) is a critical enabler of intelligent asset management in Industry 4.0, yet many existing frameworks remain difficult to operationalize due to methodological fragmentation. Common limitations include sacrificing temporal realism and class granularity for computational expediency, decoupling labeling strategy design from model hyperparameter optimization, and insufficient support for reproducibility and deployment traceability; particularly in rare-failure regimes. To address these challenges, we propose a unified, end-to-end, and fully traceable PdM framework that jointly optimizes labeling and model parameters while enforcing strict temporal fidelity. The proposed pipeline co-optimizes the failure lookahead window (\(\tau\)) and LightGBM hyperparameters within a single Bayesian optimization space using Optuna with a Tree-structured Parzen Estimator and MedianPruner, eliminating the suboptimality of fixed or decoupled labeling designs. Temporal leakage is rigorously prevented through forward-chaining cross-validation and a strictly disjoint temporal holdout evaluation. The framework is evaluated on the widely adopted Fidan synthetic dataset (876,100 samples, five classes, \(\approx 0.09\%\) failure rate), achieving state-of-the-art performance with a macro-F1 score of 0.9875, balanced accuracy of 0.9915, and superior PR-AUC compared to prior benchmarks. Computational analysis and ablation studies confirm both scalability and the non-redundant contribution of key design choices. Crucially, the pipeline exports versioned, reproducible artifacts (models, preprocessors, configurations) designed to support enterprise integration pending site-specific validation. Operational impact projections (e.g., reduced downtime, fewer dispatches) are derived from analogous deployments and require field validation for quantification.
