ABSTRACT

Rotating assets—induction motors, pumps, gearboxes, and rolling bearings—operate in acoustically and electromagnetically harsh environments, where sensor noise, non-stationarity, and class imbalance degrade fault classification. We present a deployment-oriented comparative study of Support Vector Machines (SVM), Random Forests (RF), and XGBoost for predictive maintenance (PdM) on vibration and motor-current signals when labels are scarce and sensors are noisy. The pipeline integrates (i) noise-aware feature engineering (time statistics, crest factor, spectral kurtosis–guided envelope bands, cyclostationary indicators, order-tracked spectra, and motor-current signature features), (ii) noise countermeasures (Welch PSD averaging, wavelet shrinkage, robust scaling), and (iii) imbalance handling (cost-sensitive learning, SMOTE/Borderline-SMOTE with Tomek cleaning and threshold calibration). Evaluated on two public corpora (CWRU, PHM’12) and an anonymized plant dataset, XGBoost with class weighting or SMOTE+Tomek consistently yields the highest PR-AUC and MCC, with RF a close second and minimal tuning; SVM (RBF) is competitive when band-limited features are used but is more sensitive to extreme skew and noise. Ablations show spectral-kurtosis–guided envelope features and order-tracked sideband indices contribute most at low SNR, and PR-centric thresholding improves minority-class F1 by 3–6 points over default cutoffs. A compact 30–40 feature subset preserves >98% of full-model PR-AUC while cutting inference latency ?20%, enabling cabinet-edge deployment. We discuss drift monitoring, site-specific recalibration, and CMMS integration.