Outlier Detection for Distributed Pressure Measurements

In aerodynamic modeling, wind-tunnel and flight tests are usually assumed to be the ground truth for validating and calibrating physical or data-driven models. However, the measurements, such as pressure data, can contain several outliers, which can deteriorate the calibrated models. So far, outliers have been identified by visually inspecting the data, which can be time-consuming. Hence, it is desirable to automatically detect outliers during testing to exclude them from live monitoring, identify leakages in the pressure tubing, and obtain reliable data sets for calibration involving minimal manual interaction. This work introduces two methods for this task, with semisupervised anomaly detection approaches using simulation data for learning the normal behavior. The first method is based on a proper orthogonal decomposition of the training data, whereas the second relies on a variational autoencoder. These methods are applied to wind-tunnel tests of a two-dimensional airfoil and the NASA common research model. Both methods successfully detect outliers, whereas the proper-orthogonal-decomposition- based method better classifies them. However, the methods misclassify measurement points where Reynolds-averaged Navier–Stokes features systematic errors, thus illustrating the necessity of accurate simulations for outlier detection.