<p>Rolling-element bearing faults are a major cause of induction motor failures, necessitating reliable fault diagnosis together with timely supervisory control for predictive maintenance. This paper presents a Hybrid DT framework that integrates a leak-aware Conv–BiLSTM diagnostic model with a closed-loop Sensing–Analytics–Decision–Actuation (SADA) supervisory architecture for vibration-based bearing health management using 12-kHz Case Western Reserve University (CWRU) drive-end signals. To ensure deployment-oriented performance assessment, grouped hold-out validation at the file level, training-only normalization, class balancing, and data augmentation are employed to eliminate data leakage and improve model robustness. The proposed Conv–BiLSTM classifies healthy, inner-race, outer-race, and ball fault conditions, achieving an average validation accuracy of 91.17 ± 4.02% across five random seeds, representative single-run accuracy above 93%, and a healthy-versus-faulty AUC of approximately 0.99. Beyond fault classification, the trained model is embedded within a closed-loop Hybrid DT that converts probabilistic diagnostic outputs into deterministic supervisory actions, including severity-aware torque derating, uncertainty-aware operation, and emergency shutdown. In addition to demonstrating how deep-learning diagnostics can be combined with supervisory control to provide explainable and deployment-oriented predictive maintenance, the suggested architecture creates a repeatable, leak-aware benchmark.</p>