Improved Anomaly Detection in Experimental Wind Tunnel Data using PCA
Aaron Defreitas, William Nathan Alexander, William John Devenport, Sierra N. Merkes, Scotland C. Leman, Eric P. Smith, Aurélien Borgoltz · AIAA Scitech 2020 Forum · 2020
Conformal mapping is shown to improve anomaly detection with Principal Component Analysis (PCA) for wind tunnel measurements. Experimental data from tests evaluating the performance of a series of two-dimensional airfoils are compiled. These data include multiple sensor system types and various conditions to generate a historical data set. Proper selection and interpolation to a standard distribution of sensors is shown to be critical in compiling the historical data for this PCA anomaly detection scheme. Principals of ideal flow aerodynamics and conformal mapping are used to reduce airfoil geometry dependent deterministic variance in airfoil surface pressure measurements. An eigendecomposition of the covariance between all sensor measurements are used to determine principal components that separate expected sensor covariance from unexpected covariance and noise. This reduction of deterministic variance improves the ability of the PCA scheme to detect anomalies.