Evaluating Bootstrap Resampling and SMOTE-Style Interpolation for Preserving Correlation Structure in Synthetic Airfoil Self-Noise Data

Abstract

Experimental aeroacoustic data can be costly and time-consuming to obtain, which motivates careful study of data-augmentation methods. This exploratory study compares bootstrap resampling with a SMOTE-style nearest-neighbor interpolation applied to continuous records from the NASA/UCI Airfoil Self-Noise dataset. One random 10% base sample (n = 150) was used to generate 1,353 observations by each method. Fidelity was evaluated using Pearson correlations for all 15 variable pairs, absolute correlation errors, approximate confidence intervals, correlation-difference heatmaps, standardized quantile profiles, sound-pressure-level kernel density estimates, and a multivariate energy-distance measure. Mean absolute correlation error was 0.0304 for the base sample, 0.0297 for bootstrap, and 0.0355 for SMOTE-style interpolation; maximum errors were 0.0863, 0.0543, and 0.0962, respectively. These findings describe one dataset and one sampled subset. Because the generated observations were derived from the same base sample, they are not independent experimental units, and the nominal correlationcomparison tests cannot be interpreted as independent-sample confirmation. The study therefore does not establish preservation of physical properties, improved nonlinear learning, or improved predictive performance. It provides a dataset-specific descriptive comparison that motivates repeated-sample, physically validated, and prediction-based evaluation in future work.