Explainable and Prescriptive AI for Cyber-Physical NVH Quality Control in Smart Manufacturing: A Critical Review and Research Framework

<jats:p>Artificial intelligence is increasingly used to predict noise, vibration, and harshness (NVH), detect faults, and identify quality deviations in rotating machinery and smart manufacturing. Yet high predictive accuracy does not answer the engineering questions required for process improvement: which manufacturing or geometric variable generated the response, whether the relationship is causal, what controllable setting should change, whether the change is physically and technologically feasible, and whether the intervention improves the measured product. This critical review integrates five evidence streams: direct gear and drivetrain NVH studies; explainable rotating-machinery diagnosis; explainable manufacturing-quality prediction; causal, counterfactual, and recourse methods; and digital-twin or closed-loop industrial control. The literature shows strong progress in feature attribution, saliency analysis, physics-informed representation, and interpretable fault-state estimation, but direct gear/NVH studies remain concentrated at prediction and explanation levels. To separate algorithmic sophistication from demonstrated engineering capability, we propose a cumulative Level 0–6 maturity framework spanning prediction, global explanation, local explanation, engineering root cause, actionable recommendation, physics-constrained prescription, and closed-loop validation. The framework supports comparison of studies and design of cyber-physical NVH quality-control systems that progress from explanation to traceable, feasible, and experimentally validated manufacturing action.</jats:p>