Lightweight Cross-Condition Rolling Bearing Fault Diagnosis via Time-Domain and Multi-Order Feature Fusion

Abstract

Changes in speed, torque, and load alter bearing-vibration distributions and weaken models trained under fixed conditions. This paper combines twelve time-domain statistics with fifteen mechanically defined envelope-order energy ratios and classifies the resulting 27-dimensional vector using a random forest. Strict condition-wise holdout is used instead of random window splitting. On CWRU, the method achieves 98.10% Macro-F1 under leave-one-load-out testing, 90.87% under 5-dB additive white Gaussian noise, and $88.87 \%$ under a 5-dB composite interference containing shaft harmonics, random decaying impacts, and Gaussian background. Under the same composite interference, the time-domain baseline obtains 79.17%. On Paderborn, leave-one-condition-out Macro-F1 improves from $\mathbf{6 7 . 9 7 \%}$ to $\mathbf{8 0 . 2 9 \%}$. Parameter-grid and leave-one-order-feature-out analyses further show a broad stable operating region and identify the most influential physical orders. The method therefore provides a compact and interpretable accuracy–robustness compromise, although large speed extrapolation remains difficult.