Optimal Feature Complexity for Small-Sample Bearing Fault Detection in Manufacturing

Abstract Manufacturing environments increasingly demand cost-effective bearing fault detection solutions that balance diagnostic accuracy with implementation practicality. This study investigates optimal feature complexity levels for small-sample bearing condition monitoring using the standardized Paderborn University benchmark dataset. Through systematic evaluation of 15 bearing specimens across three complexity categories, we analyze the performance-cost trade-offs critical for manufacturing deployment. Results demonstrate that moderate-complexity features achieve superior discriminative power, with spectral centroid reaching Cohen’s d=2.06 compared to fractal dimension’s d=0.40. Machine learning methods significantly outperform traditional threshold approaches, with Random Forest achieving 80% accuracy versus 40% for RMS-based detection under small-sample constraints typical in manufacturing. Statistical validation through leave-one-out cross-validation reveals that moderate complexity provides optimal performance-per-implementation-cost ratios. The study establishes practical guidelines for manufacturing engineers, demonstrating that strategic feature selection outweighs algorithmic sophistication for industrial bearing diagnostics under realistic data constraints.