An Uncertainty-Aware Quantum–Classical Digital Twin Framework for Predictive Maintenance of Rotating Machinery

Digital twins are increasingly used to monitor rotating machinery, and a growing literature proposes adding quantum machine-learning components to them. Whether a near-term quantum model improves a maintenance twin, and how it should be integrated so that it cannot degrade the twin when it does not, has not been tested in a controlled setting. This paper presents a framework in which a six-qubit variational quantum classifier (VQC) ensemble and a classical multilayer-perceptron ensemble are combined through an uncertainty-aware gate: calibrated predictive entropy decides whether the quantum circuit is executed at all (an escalation cascade) and how its output is weighted (entropy-weighted fusion). The twin state (speed, bearing temperature, defect severity and its growth rate) is tracked by an extended Kalman filter that is synchronized with the edge node only when the normalized innovation exceeds a threshold, and the severity belief drives a remaining-useful-life (RUL) estimate and a cost-based maintenance rule. A physics-based motor–pump–bearing simulator with McFadden–Smith defect impulses, motor-current sidebands and fault injection provides 2,500 labelled windows over 0–20 dB and 80 run-to-failure trajectories. Two results are reported without softening. First, the gate keeps the hybrid at the accuracy of the better model (98.0 % against 98.2 % for the classical ensemble and 88.8 % for the VQC) while executing the circuit on 4.4 % of windows, and holds 98.0 % when simulated hardware noise drives the VQC alone to 61.4 %. Second, the VQC did not outperform a parameter-matched 61-parameter classical model under any tested condition, so no quantum advantage is claimed. Innovation-triggered synchronization reduced uplink traffic by 87 % at a 13 % increase in severity-tracking error, and the twin-based maintenance rule matched the best oracle-tuned periodic interval without knowledge of the fleet life distribution. All code, data generators and results are released for reproduction.