Predictive Maintenance (PdM) in Industry 4.0 relies on prognostics and health management (PHM) frameworks that combine vibration analysis with AI, including machine learning (ML) and lightweight deep learning (DL). A physics-aware workflow engineer’s time-domain features, applies dimensionality reduction, and trains compact learners for fast inference. Using the HUST bearing dataset under a strict subject-out split (all preprocessing fit on the training set only), a LightGBM classifier attains 98.01% test accuracy and a macro-AUC of ≈0.999, indicating robust class discrimination. The resulting model is small and efficient, enabling integration into cyber-physical systems (CPS) on modest hardware. Practical steps for PdM deployment: edge-side feature extraction, stable normalization learned offline, and periodic model refresh to track operating-point drift. The approach emphasizes reproducibility with transparent splits and no leakage. Limitations include the laboratory setting, wire-cut incipient defects, a single accelerometer channel, and restricted operating conditions; Future work targets prognostics by deriving a monotone health index and learning degradation trajectories to produce explainable RUL estimates within PHM. The study shows that a physics-aware ML pipeline delivers reliable bearing fault diagnosis while meeting PdM constraints on latency and compute.