Data‑Driven Analysis of Gear Manufacturing Tolerances on Transmission Error and Noise Excitation: A Machine Learning Perspective

Transmission error (TE) is one of the most important sources of gear noise and vibration. Manufacturing tolerances and assembly shifts introduce deviations in tooth geometry that produce periodic mesh disturbances; these disturbances excite the drivetrain and radiate as airborne sound. While many studies have modelled individual tolerance effects in deterministic simulations, few have evaluated the combined influence of realistic tolerance distributions using open Monte–Carlo data. This work analyses the public Gear Statistical tolerance analysis dataset (≈40 k samples) to answer the following questions: (Q1) How accurately can the kinematic transmission error be predicted from measured tolerance and process‑shift data? (Q2) Which individual tolerances and their interactions have the greatest impact on TE? (Q3) Can a TE‑derived proxy quantify noise‑critical excitation without an acoustic model? (Q4) What tolerance combinations minimise TE and noise while respecting manufacturing cost? (Q5) Is there a quantifiable trade‑off between tolerance tightening cost and noise reduction? To address these questions we formulate hypotheses: H1 Non‑linear machine‑learning models achieve high predictive accuracy for TE (R² > 0.85) compared with linear baselines. H2 A small subset of tolerances—profile and lead errors—dominate the TE variance and interact non‑linearly. H3 A relative noise proxy (RNP) derived from TE faithfully ranks noise‑critical excitation across tolerance combinations. H4 A Pareto front exists in the cost–noise plane, enabling cost‑effective tolerance optimisation. H5 Targeted adjustment of the top two tolerances yields larger noise‑reduction per cost than uniform tightening. The following sections describe the data, the modelling framework and the results that support these hypotheses.