We propose a hierarchical Bayesian digital twin framework for real-time predictive maintenance of rotating machinery. Our approach overcomes the lack of formal reliability guarantees and multimodal fusion in existing systems. The raw vibration and temperature streams are abstracted into three interpretable layers: time-domain statistics, multiscale temporal trends with inter-sensor correlations, and a unified health index. Each layer has conjugate priors that allow for closed-form posterior updates when new data is received. Our solution achieves sub-millisecond per-update latencies on commodity hardware without relying on computationally intensive physics solvers. Diagnostic fidelity is also maintained. On a benchmark dataset of 10,000 h of vibration and temperature recordings, the fused model achieves a true positive rate of 92.3% and a false positive rate of 2.5% under a user-specified α = 0.05 threshold. The area under the ROC curve is 0.96. We demonstrate that the chosen α limits the system-wide false alarm likelihood, providing maintenance planners with principled control over the early warning trade-off. This combination of interpretability, computational efficiency, and probabilistic rigor paves the way for scalable, dependable monitoring systems capable of adapting to new sensor modalities and operating situations.