Physics-Informed Order-Topology Analysis for Source-Gear Inference of Manufacturing Deviations in Electric-Vehicle Reducer Gears Under Variable-Speed Operation

Manufacturing-induced tooth-surface deviations in electric-vehicle reducer gears can generate tonal or impulsive responses during speed sweeps. This study evaluates a physics-informed order-topology framework for identifying five manufacturing classes, inferring the source gear, and distinguishing deviation morphology from one vibration channel and a synchronous speed reference. Computed order tracking and prespecified features describe nominal mesh orders, dataset-specific 40th- and 62nd-order pinion ghost families, time-domain impulsiveness, and a ratio-parametric Output-Gear Harmonic Comb Alignment Index (OG-HCAI). A 107-record development subset was used to freeze processing and models. The primary file-disjoint confirmation comprised 86 recordings from 51 category–measurement-ID groups; a stricter post hoc category–measurement-ID-disjoint sensitivity analysis retained 83 recordings from 48 groups without refitting. Primary balanced accuracies were 0.923 for five-class Random Forest classification, 0.937 for Logistic Regression source-gear inference, and 0.935 for morphology; the sensitivity-analysis values were 0.940, 0.950, and 0.956. Frozen indicator areas under the receiver operating characteristic curve (ROC AUCs) were 0.902, 0.991, 0.885, and 1.000 for the 40th order, 62nd order, bump kurtosis, and OG-HCAI, respectively, with the last result limited to 14 records from seven positive groups. The results support interpretable class-level diagnosis while requiring broader cross-reducer validation.